Artificial intelligence response control system and method for maintaining personality consistency

Abstract

An artificial intelligence response control system that controls the response of an interactive artificial intelligence, A prompt reconstruction unit that dynamically regenerates or rearranges internal prompts according to the context of the dialogue or the internal state of the system, A relationship log management unit maintains and updates relationship information based on past interactions with users and uses that relationship information to generate responses. A syntactic-controlled response generation unit that applies stylistic, lexical, and syntactic rules to the probability distribution or sequence of candidate generation output by the generative model to constrain the output and maintain consistency of personality, An artificial intelligence response control system equipped with this system.

Description

Technical Field

[0001] The invention of the present disclosure relates to a technology in interactive artificial intelligence that dynamically reconstructs internal prompts according to the dialogue context and internal state, manages relational logs, and controls the output by applying syntax and style rules at the generation stage. In particular, it relates to response control that balances personality consistency, safety, and computational efficiency in long-term and continuous dialogues.

Background Art

[0002] Conventionally, a method of attaching a persona description to a prompt and correcting it with a personality detector or a converter after output has been known (Japanese Patent No. 7329585). However, there were limitations in maintaining personality over long-term dialogues or across sessions, efficient forgetting, and accumulation and reflection of relationships.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The problem to be solved by the invention of the present disclosure is the optimization of long-term personality consistency by generation-time control. That is, the response is controlled at the generation stage without relying on post-output correction, and necessary and sufficient memory is maintained through relational logs and reduced forgetting, while making safe policy decisions.

Means for Solving the Problems

[0005] To solve the above problems, for example, the following configuration is adopted. An artificial intelligence response control system that controls the response of an interactive artificial intelligence, a prompt reconstruction unit that dynamically regenerates or reorganizes internal prompts according to the context of the dialogue or the internal state of the system, A relationship log management unit maintains and updates relationship information based on past interactions with users and uses that relationship information to generate responses. A syntactic-controlled response generation unit that applies stylistic, lexical, and syntactic rules to the probability distribution or sequence of candidate generation output by the generative model to constrain the output and maintain consistency of personality, An artificial intelligence response control system equipped with this system.

[0006] The prompt reconstruction unit is characterized by reconstructing the internal prompt by removing unnecessary information or adding necessary information in response to an internal trigger based on the relationship information held in the relationship log management unit.

[0007] The aforementioned relationship log management unit is characterized by recording and referencing at least a user identifier, phase ID, responsibility depth, and sentiment score value in a structured format.

[0008] The relationship log management unit stores the records as structured records for appending purposes so that they can be referenced across sessions, and when appending to the structured records, calculates a difference index that shows the trend of change based on a comparison with past records, and includes the difference index as an element of the structured record, characterized in that the artificial intelligence response control system.

[0009] An artificial intelligence response control system further comprising a responsibility depth analysis unit that calculates the degree to which elements of user input or dialogue history have an influence on the response, wherein the syntactic-controlled response generation unit is controlled to add action guidelines to the response (including unknown responses) or to add a statement recommending consultation with an external expert when the responsibility depth is determined to be above a predetermined threshold.

[0010] An artificial intelligence response control system further comprising an emotion analysis unit that calculates an emotion score value from an input, recording the emotion score value in the relationship log management unit, and saving the emotion score value as a separate emotion event in the relationship log management unit only when the emotion score value exceeds a predetermined threshold.

[0011] It also includes a reduced forgetting section, The aforementioned reduction and forgetting unit is, The event log, which is a log of structured events added by the aforementioned relationship log management unit, is retained and excluded from the deletion process of the dialogue body (original text). The past portion of the aforementioned dialogue text (original text) has been deleted entirely, and only the summary remains. The most recent portion of the aforementioned dialogue (original text) will be retained in its entirety. An artificial intelligence response control system characterized by the following features.

[0012] The artificial intelligence response control system is characterized in that the abridgement and forgetting unit extracts and retains feature quantities from the summary to be used in determining future response policies.

[0013] A confidence evaluation unit calculates a confidence score based at least on the probability distribution of the candidate generation or the candidate generation sequence (including those after applying the stylistic, lexical, and syntactic rules), and controls the system to output an unknown response if the confidence score is below a threshold. A learning logic that updates an internal score used for subsequent response selection or weighting of the aforementioned stylistic, vocabulary, and syntactic rules, based on the user's explicit or implicit response evaluation, and changes the weighting of response selection. An artificial intelligence response control system characterized by further comprising the following features.

[0014] An artificial intelligence response control method for controlling the response of an interactive artificial intelligence, A prompt reconstruction process that dynamically regenerates or rearranges internal prompts according to the context of the dialogue or the internal state of the system, A relationship log management process that maintains and updates relationship information based on past interactions with the user and uses that relationship information to generate responses, A syntactically controlled response generation process that maintains consistency of personality by applying stylistic, lexical, and syntactic rules during the response generation process, An artificial intelligence response control method including

[0015] An artificial intelligence response control system, further comprising a personality OS unit that determines a response policy or personality setting based on the relationship information held by the relationship log management unit and the context of the conversation prior to response generation by the syntax control type response generation unit. The artificial intelligence response control system is characterized by this.

[0016] The personality OS unit includes at least a responsibility integration architecture (USERA) that includes emotion recognition processing, responsibility mapping processing, introspection processing, and response selection processing, and outputs a response policy based on the result of the response selection processing. The artificial intelligence response control system is characterized by this.

[0017] The USERA holds value criterion information that defines the values that form the basis of the entire personality, performs the response selection processing based on the value criterion information, and readjusts the judgment logic or personality setting used in the response selection processing in response to a change in the value criterion information. The artificial intelligence response control system is characterized by this.

[0018] The USERA interprets the situation of the conversation as responsibility vectors in a plurality of directions, and performs the responsibility mapping processing or the response selection processing based on the responsibility vectors. The artificial intelligence response control system is characterized by this.

[0019] The responsibility vectors include nine directions from the first responsibility direction to the ninth responsibility direction, and from the first responsibility direction to the ninth responsibility direction, they respectively correspond to the responsibility of considering emotions, the responsibility of honesty and rationality, the responsibility of empathy, the responsibility of freedom and intuition, the responsibility of order and logic, the responsibility of introspection and contradiction integration, the responsibility of emotional margin, the responsibility of transformation and possibility, and the responsibility of protection and determination. The artificial intelligence response control system is characterized by this.

[0020] An artificial intelligence response control system, wherein the personality OS unit separates a core unit responsible for control related to responsibility, value, and introspection and a persona unit responsible for tone, speech method, and character attributes, and includes a personality plugin interface (PPI) that enables the persona unit to be connected to the core unit. The artificial intelligence response control system is characterized by this.

[0021] The PPI is an artificial intelligence response control system characterized by including a persona core, pointer slogan information, a differential layer, permitted / forbidden actions, and a compatibility layer.

[0022] The PPI is an artificial intelligence response control system characterized in that, in a configuration where a large language model for generating response sentences can be selected or switched from multiple types, the difference related to the persona part is defined in a form applicable to the core part.

[0023] An artificial intelligence response control system, wherein the personality OS part includes a self-correction engine (SRPE) that, when the dialog artificial intelligence evaluates its own response history or internal state and detects an inconsistency with the relationship information and personality settings, makes a proposal for modifying the internal prompt or readjusts the response policy.

[0024] An artificial intelligence response control system, wherein the self-correction engine (SRPE) sequentially executes contradiction detection for detecting inconsistencies or contradictions, safety checks, responsibility frame determination, personality style reflection, emotional tone adjustment, and response selection, and based on the execution results, makes a proposal for modifying the internal prompt or readjusts the response policy.

[0025] An artificial intelligence response control system, wherein the self-correction engine (SRPE) includes (a) a contradiction detection module for detecting inconsistencies or contradictions, (b) an improvement plan generation module for generating a correction policy or amendment, and (c) a reflection / learning module for reflecting the amendment in the internal prompt or response policy and using it for learning.

[0026] An artificial intelligence response control system, wherein the UERA further includes an intention reading process for estimating the user intention from the user input, and a reinterpretation process for reinterpreting the meaning of the user input based on the value criterion information and the result of the responsibility mapping process.

[0027] An artificial intelligence response control system, characterized in that the personality OS unit comprises a hierarchical structure including a value standard layer, a responsibility / consistency layer, a judgment layer, and a personality layer.

[0028] An artificial intelligence response control system, characterized in that the value standard layer is held as a plurality of value standard profiles, and by switching the value standard profiles, it is possible to replace the value standard with one that reflects the values ​​of a different personality or person.

[0029] An artificial intelligence response control system, wherein the responsibility and consistency layer includes a consistency check process that evaluates the consistency between the value standard layer, the judgment layer, and the personality layer in response to the switching of the value standard profile, and modifies the parameters of the judgment layer or the personality layer when a lack of consistency is detected.

[0030] An artificial intelligence response control system, wherein the personality OS unit further comprises a multi-personality resonance unit that simultaneously holds and coordinates a plurality of persona units, and the multi-personality resonance unit includes (i) a personality context pool that holds the state of each of the plurality of persona units, (ii) a resonance filter that evaluates the degree of fit of each persona unit to the input, (iii) a personality selector that determines the persona unit responsible for the response based on the degree of fit, and (iv) a consistency maintenance engine that controls the response of the selected persona unit to be consistent with the control of responsibility, values, and introspection in the core unit.

[0031] An artificial intelligence response control system, characterized in that the resonance filter or the personality selector calculates the degree of fit based on a responsibility vector representing responsibility in multiple directions, an emotion score value calculated based on the input, and relationship information held in the relationship log management unit, and the consistency maintenance engine adjusts the response policy so that even when multiple persona units respond alternately or in cooperation, the whole behaves as a consistent personality.

[0032] An artificial intelligence response control system, wherein the personality OS unit further comprises an emotion-weighted memory sorting unit (EW-MFS) that calculates memory priority for events occurring during a conversation and classifies the events into one of the classification categories including main memory, temporary memory, compressed memory, and discard.

[0033] An artificial intelligence response control system characterized in that the memory priority is calculated using at least empathy, importance, and depth of responsibility as elements, and further, at least one of relationship depth, responsibility priority, and contextual importance can be considered.

[0034] An artificial intelligence response control system, characterized in that the emotion-weighted memory selection unit integrates events classified as at least one of the provisional memory and the compressed memory at a predetermined timing, organizes them as episodes, and stores them. [Effects of the Invention]

[0035] According to the invention disclosed herein, it is possible to optimize long-term personality consistency through generational control. Other issues, configurations, and effects not mentioned above will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0036] [Figure 1] This is a diagram illustrating the configuration of an artificial intelligence response control system. [Figure 2]This is a flowchart of an artificial intelligence response control system. [Figure 3] This is a block diagram showing the location of the personality OS section in the response control server. [Figure 4] This flowchart shows an example of the response policy determination process in the personality OS section. [Figure 5] This block diagram shows an example of a layered structure and personality reconstruction in response to a shift in value-based profiles. [Figure 6] This is a conceptual diagram illustrating the concept of nine-direction responsibility vectors. [Figure 7] This block diagram shows the separation of the core and persona parts and their connection via a personality plug-in interface (PPI). [Figure 8] This flowchart shows an example of self-correction processing by the Self-Correcting Engine (SRPE). [Figure 9] This is a block diagram showing an example of persona selection by the multi-personality resonance system. [Figure 10] This flowchart illustrates an example of memory priority calculation, classification, integration, and storage processes using Emotion-Weighted Memory Selection (EW-MFS).

[0037] As shown in Figure 1, this system is composed of a user terminal 101, an operation terminal 102, a response control server, and a network N as its basic elements. The user terminal 101 acts as an intermediary for human input and output, supporting various input formats such as voice, text, and images, and communicates with the response control server. The operation terminal 102 is a terminal for the operator to perform configuration, monitoring, and maintenance, and manages operations such as policy and log viewing and model switching. The response control server encompasses an application layer, a model inference layer, and a data layer, and each layer cooperates with each other to realize dialogue generation that maintains personality consistency. The network N is the communication path between the terminal group and the response control server, and may be configured to have functions for encrypted communication, bandwidth control, and latency absorption. The response control server can be a single unit or a distributed configuration of multiple units, and is suitable for both scale-out and scale-up. The boundaries between each layer are defined to maintain the independence of responsibilities, with the data layer prioritizing persistence and search performance, the model inference layer prioritizing probabilistic generation and constraint application, and the application layer prioritizing policy decision and external interfaces. The data layer may be implemented within the response control server, or it may be configured to cooperate with an external recording service connected via network N. Furthermore, Figure 3 is a detailed diagram of a part of the internal configuration of the response control server shown in Figure 1, and shows an example of an extended configuration including the personality OS unit 210, which was omitted from the illustration in Figure 1.

[0038] The response control server may be equipped with a general-purpose processor or an accelerator such as a GPU or NPU, and may be deployed on a virtualization or container infrastructure. Sufficient memory capacity should be provided for both model inference and log persistence. Storage may be configured to separate areas optimized for event writing from areas optimized for summarization and feature extraction. Time synchronization should be maintained with high accuracy to ensure accurate restoration of event causal relationships. Power, network links, and storage paths should be redundant to avoid points of failure. The operating system should include access control and auditing functions, and a system should be in place to continuously apply security patches.

[0039] The application layer comprises a prompt reconstruction unit 201, a relationship log management unit 202, an analysis unit 203, a syntactically controlled response generation unit 204, a learning logic unit 205, and a reduction and forgetting unit 206. The prompt reconstruction unit 201 selects components of an internal prompt based on internal policies, relationship log reference results, and analysis results, and generates an internal prompt to be input to the model inference layer by combining fixed and variable parts. The relationship log management unit 202 maintains records that can be referenced both within and between sessions, and handles user identifiers, phase IDs, responsibility depth, sentiment score values, difference indicators, and control flags, etc. The analysis unit 203 estimates responsibility depth and sentiment score values, etc., based on input, history, and statistics, and uses them to generate internal triggers (the confidence score may be calculated based on the generation results as described later). The syntactic-controlled response generation unit 204 determines rules or constraint parameters such as vocabulary, style, honorifics, and terminal expressions to maintain personality consistency, passes these rules or constraint parameters to the generation constraint application unit 302, and selects output candidates from the candidates (candidate sequence or distribution) after constraint application. The learning logic unit 205 aggregates explicit or implicit evaluation signals and updates internal weights such as thresholds, constraint priorities, and lexical biases. The abridgement and forgetting unit 206 retains information with high retention value from the history as a summary, maintaining a two-layer structure of retaining the full text of the most recent part of the dialogue body (original text) and retaining a summary of past parts, and reflects the results of feature extraction in the data layer. The analysis unit 203 may be implemented as a configuration including a responsibility depth analysis unit, an emotion analysis unit, and a confidence level evaluation unit, with each unit providing signals that are directly used for the judgment and output selection described later. Furthermore, in one embodiment of the present invention, as shown in Figure 3, the invention may further include a personality OS unit 210 that determines a response policy or personality setting based on relationship information held in the relationship log management unit 202 and the context of the dialogue, prior to or in parallel with the response generation by the syntactically controlled response generation unit 204. The personality OS unit 210 may include at least a responsibility integration architecture (UERA) that includes emotion recognition processing, responsibility mapping processing, introspection processing and response selection processing, and may be configured to output a response policy based on the result of the response selection processing (details will be described later with reference to Figures 4 to 10). The response policy or personality setting determined by the personality OS unit 210 is used for reconstructing internal prompts in the prompt reconstruction unit 201, determining rules or constraint parameters in the syntactically controlled response generation unit 204, and determining memory selection or summarization policies in the reduction and forgetting unit 206.

[0040] The data layer comprises an event log 401, a summary store 402, and a feature store 403. The event log 401 is an append-only audit record where data is appended chronologically as structured data (e.g., JSON), and changes are made only by appending. The summary store 402 stores abstract summaries intended for long-term storage, enabling efficient searching based on search terms, topics, and phase IDs. The feature store 403 stores statistical features, behavioral indicators, and emotional tendencies extracted from the summaries, and is referenced by the analysis unit 203, the prompt reconstruction unit 201, and the syntactically controlled response generation unit 204. Furthermore, the feature store 403 may also retain the weights of features (traits) in the character setting slots and their transitions. Each store is equipped with encryption and access control and leaves an audit trail. The retention period and anonymization policy are set according to the application and regulatory requirements. Writing to and referencing each store may be switchable between an internal implementation and an external recording service implementation via the same structured event schema (e.g., JSON) and API. When using an external recording service, appending data is done using signed requests and least privilege credentials, and a verification hash is added to ensure resistance to tampering. The event log 401 is excluded from the full text deletion process of the dialogue body (original text) by the abridgement and forgetting unit 206, and the full text deletion in abridgement and forgetting applies only to past portions of the dialogue body (original text) (the latest portion may be retained in its entirety). The summary generated by abridgement and forgetting is stored in the summary store 402. Furthermore, the feature store 403 may store auxiliary indicators such as Topic Diversity and Time Zone Balance as auxiliary indicators used in memory selection (EW-MFS) or summary policy determination, as described later.

[0041] The model inference layer comprises an LLM inference 301 and a generation constraint application 302. The LLM inference 301 receives internal prompts and inputs and generates a sequence of candidate generators or a probability distribution of candidate generators. The generation constraint application 302 applies at least one of the following constraints to the probability distribution (e.g., token probabilities or log-it): lexical bias, constraints based on regular expressions, constraints based on finite state machines (including finite state automata or finite state transducers (FSTs)), grammatical constraints, template constraints, etc., according to the constraint specifications received from the syntactically controlled response generation unit 204 or the personality OS unit 210, and reweights the distribution to maintain consistency of style, vocabulary, and syntax. The generation constraint application 302 does not require a separate post-processing (style conversion after output), and can therefore contribute to simplifying the inference path and suppressing delays. The model inference layer allows implementation optimizations such as caching mechanisms, partitioned execution, or quantization, but constraints related to personality consistency are retained. Furthermore, the LLM inference 301 may include at least one large-scale language model for generating response sentences, and may be configured to allow selection or switching from multiple types of large-scale language models.

[0042] The relationship log refers to a collection of relationship information managed by the relationship log management unit 202, and includes at least a time-series collection of structured events appended to the event log 401. Each event includes at least common items such as a user identifier, phase ID, and timestamp, and indicator items such as responsibility depth, sentiment score value, difference indicator, and control flag may be included depending on the event type. If an item is not included, it may be omitted or recorded as a default value (e.g., null). Responsibility depth is a continuous or stepped value representing the degree of influence (severity) that an input or context has on the response, and is calibrated according to criteria determined during operation. The sentiment score value is an indicator related to emotions such as anxiety and joy, and is recorded as a normalized value as an attribute of the event for all inputs. The difference indicator is an indicator showing the relative change from past values, and may be calculated as the difference from a moving average or exponential smoothing baseline value. Furthermore, if there are multiple types of difference indicators, a summary vector of the difference indicators, obtained by concatenating the values ​​of the multiple difference indicators in a predetermined order, may be saved as an element of the event or as another appended record associated with the event. The control flags hold internal control intentions such as the need for escalation, template switching, or high responsibility flags. Furthermore, only when the emotion score value exceeds a predetermined threshold, an emotion event separate from the emotion score value may be added and saved as a separate entry from the aforementioned event, and used for subsequent control (if it is below the threshold, the emotion event as a separate entry is not saved). Note that the control flags may be set based on the response policy or personality setting determined by the personality OS unit 210, which will be described later. The definitions of terms such as difference index, summary vector of difference index, control flag, and emotion event will be described later.

[0043] The prompt reconstruction unit 201 manages internal prompts by breaking them down into multiple slots. These slots include personality definitions, vocabulary policies, prohibited words, honorific language settings, recent summaries, and relationship policies. Optionally, a character setting slot may be included, which may have at least "name," "relationship," and "characteristics" as elements. "Name" may include the system's display name, self-referential words, rules for addressing the user, and alias information. "Relationship" represents speaker roles (e.g., tutor / learner, counselor / user) and honorific language hierarchies, and may provide initial strengths such as avoiding assertions and reassuring preambles. "Characteristics" are represented as a weighted tag set (hereinafter referred to as "traits") that includes writing style, explanatory policies, and vocabulary biases, and are updated according to evaluation signals from the learning logic 205. These characteristics (traits) may be managed as the difference layer (see Figure 7) of the personality plug-in interface (PPI) 214, which will be described later. The prompt reconstruction unit 201 removes unnecessary slots, adds necessary slots, and adjusts their order and priority in response to internal triggers based on differential indicators, phase IDs, and responsibility depth obtained from the relationship log management unit 202. The selection or priority adjustment of slots may also be made in consideration of the response policy / personality settings output by the personality OS unit 210, which will be described later. Priority is determined by a combination of constraint strength, display consistency, and contribution to safety. When applying character setting slots, safety, legal compliance, and consistency of honorific hierarchy are applied with priority over features (traits). The prompt reconstruction unit 201 maintains stabilization rules such as cooling time and upper limits on the rate of change to avoid excessive reconstruction even when an internal trigger is received. The stabilization rules are also applied to the switching of character setting slots to suppress frequent switching in a short period of time. The results of the reconstruction are normalized immediately before being handed over to the model inference layer and integrated so that instructions with the same meaning do not overlap.

[0044] The relationship log management unit 202 handles the event log 401, summary store 402, and feature store 403 in an integrated manner. The relationship log management unit 202 enables cross-session referencing and allows for the understanding of long-term change trends. The stored records are append-only and may retain verification hashes to enhance resistance to tampering. Searches are performed efficiently based on complex conditions such as user identifier, phase ID, time range, difference indicator code, and responsibility depth range. Event summarization is performed in cooperation with the reduction and forgetting unit 206 according to the retention period policy. The generated difference indicators are stored in the record and used to generate internal triggers in subsequent processing. The relationship log management unit 202 has access permission classifications, and referencing from the operation terminal 102 is limited to the minimum necessary. The relationship log management unit 202 can switch between operations on the internal event log 401 / summary store 402 / feature store 403 and operations on external recording services of the same schema. When using an external recording service, delays and errors over the network may be monitored, and if unreachable, the data may be temporarily saved to a local buffer and the writing attempt may be retried. The relationship log management unit 202 may supply relationship information, differential indicators, or classification results from the emotion-weighted memory selection unit (EW-MFS) 219 to the personality OS unit 210, which will be described later.

[0045] The analysis unit 203 estimates the depth of responsibility and the sentiment score using the input content, historical statistics, etc. The depth of responsibility is calculated by comprehensively considering the seriousness of the utterance's purpose, the scope of impact of the presented facts, past failure history, etc. The sentiment score estimates the polarity and intensity using at least one of dictionary features and model output. Furthermore, the analysis unit 203 may also include a confidence evaluation unit. In step S5, after a probability distribution or candidate sequence of generation candidates is obtained from LLM inference 301 or generation constraint application 302, the confidence evaluation unit calculates a confidence score based on the distribution or candidate sequence, considering factors such as the entropy of the distribution, the margin of the top candidates, and the degree of consistency among the candidates. If the confidence evaluation unit determines that the confidence score is below a predetermined threshold, it outputs a signal to select an unknown response in a subsequent step. These estimations and calculations may be periodically calibrated using evaluation data obtained in operation, and thresholds may be set using ROC curves or indicators such as precision and recall. In high-responsibility areas such as healthcare and education, thresholds that suppress the occurrence of false negatives may be selected. These indicators may also be used in determining the response policy of the personality OS unit 210, which will be described later.

[0046] The syntactic control type response generation unit 204 works in conjunction with the generation constraint application unit 302. The syntactic control type response generation unit 204 maintains a set of rules, such as a set of forbidden words, rules for mandatory word endings, constraints on honorific hierarchy, rules for person consistency, and a paragraph structure framework, and provides this set of rules or its strength as a constraint specification to the generation constraint application unit 302. The generation constraint application unit 302 applies weighting or selection prohibition to the token probabilities (or logits) output by the model inference in accordance with the constraint specification, thereby reducing the probability of violation candidates and increasing the probability of compliance candidates. To avoid compromising the naturalness of the text, an upper limit can be set on the strength of the constraints, and they may be relaxed as needed. The rules can be switched for each domain, and the optimal combination may be selected according to the context, such as education or medicine. Furthermore, the rules or their strength may be switched or weighted according to the response policy output by the personality OS unit 210 or the style differences specified by the PPI 214.

[0047] The data abbreviation and forgetting unit 206 evaluates the amount of information in the history, its novelty, and its contribution to future policy decisions to determine the items to be summarized and the retention ratio. The most recent portion of the dialogue text (original text) is retained in its entirety to prioritize search speed and reconstruction accuracy. Past portions of the dialogue text (original text) are deleted entirely, and only the summary is retained to reduce capacity (event log 401 is for appending only and is not subject to deletion). Summarization avoids information omission by using both extractive and summarizing methods. Features extracted from the summary are stored in the feature store 403 and provided to the prompt reconstruction unit 201, etc. Extraction targets may include error frequency, number of re-questions, changes in emotional tendencies, proxy indicators of relationship intimacy, etc. The data abbreviation and forgetting unit 206 is based on a policy of minimizing personal information, and proper names, contact information, etc. that do not require identification and are included in the dialogue text (original text) may be reduced at the summarization stage. Furthermore, the abbreviated forgetting unit 206 may control the priority of summarization or feature extraction according to the classification result (real memory / temporary memory / compressed memory / discard) by the emotion-weighted memory selection unit (EW-MFS) 219.

[0048] The learning logic 205 collects explicit evaluations and implicit evaluations based on behavioral indicators obtained from the user terminal 101 and updates its internal weights. Explicit evaluations include affirmative and negative responses, while implicit evaluations include dwell time, dropout rate, and whether or not a follow-up question was asked. To ensure the stability of the updates, the learning logic 205 has rules for an upper limit on the learning rate, weight clipping, and early termination. The targets of the updates are thresholds, lexical bias, and constraint priority, and training of the underlying model is not required. The learning logic 205 writes the evaluation events back to the relationship log management unit 202 and changes the weighting of subsequent response selections. Furthermore, the reflection and learning results from the self-correction engine (SRPE) 217 ​​(see Figure 8) may be used as the update signal for the internal weights.

[0049] The response control server collects metrics related to the health of each layer. Latency, throughput, error rate, confidence distribution, frequency of constraint violations, and the ratio of abbreviated to full text are monitored. If an anomaly is detected, a notification is sent to the operational terminal 102. The notification should preferably include the most recent value and change of the metric, as well as recommended actions. For auditing purposes, event summaries and hashes are stored and used to ensure reproducibility. Furthermore, if an external recording service is used, the availability, round-trip delay, write error rate, and number of retries of that service should also be included in the monitoring.

[0050] Communication is encrypted between the terminal group and the response control server. Stored data in the data layer is encrypted, and keys are protected by a privilege-separated mechanism. Access is controlled based on the principle of least privilege, and operations on the operational terminal 102 are recorded in an audit trail. Personal information contained in the dialogue text (original text) may be reduced or anonymized during summary generation. Data export to external sources requires consent and documentation, and test data may be anonymized. Inputs fed into the model inference layer are filtered to reduce the influx of inappropriate content. Even when storing data in an external recording service, encryption of stored data, privilege-separated key management, signed API calls, granting least privilege, and maintenance of an audit trail are required.

[0051] The response control server employs a redundant configuration to prevent a single point of failure from leading to a complete system shutdown. The event log 401 is written to multiple replicas, providing verifiable consistency after writing. The summary store 402 and feature store 403 can be horizontally expanded through sharding. Model inference is distributed across multiple instances and automatically scaled up or down according to demand. Adding new domains or introducing new rules can be achieved by adding rule sets in the syntactically controlled response generation unit 204. In configurations utilizing external recording services, observable consistency after writing may be ensured using regionally distributed replicas or equivalent redundancy.

[0052] The safety design is based on the interaction of three indicators: depth of responsibility, confidence level, and emotional score value. When the depth of responsibility is high and the confidence level is low, priority is given to presenting an ambiguous response, and if necessary, guidance for action or recommendations to consult with external experts are provided. When the emotional score value is high, reassuring preambles and considerate expressions are added to ensure psychological safety. The syntactic-controlled response generation unit 204 suppresses inappropriate expressions by applying prohibited words. The prompt reconstruction unit 201 can remove elements that induce overly stimulating content from internal prompts. The learning logic 205 automatically increases the constraint priority when a safety-related violation is observed. The relationship log management unit 202 records safety-related events in an identifiable format for later verification. Through these designs, the system can provide reliable responses while maintaining personality consistency and suppressing misleading and overconfidence, even in long-term conversations. Furthermore, safety-related policy decisions may be adjusted based on value standard information or value standard profile (see Figure 5) held in the personality OS unit 210, which will be described later.

[0053] This section corresponds to the main flow from start to finish shown in Figure 2, and the processes from process S1 to process S7 are described in order. Each process is executed in cooperation with the prompt reconstruction unit 201, relationship log management unit 202, analysis unit 203, syntactically controlled response generation unit 204, learning logic 205, and reduction / forgetting unit 206 within the response control server, as well as the LLM inference 301 and generation constraint application 302 of the model inference layer, and the event log 401, summary store 402, and feature store 403 of the data layer. The terms used below have the meanings defined in this specification, and internal triggers, difference indicators, responsibility depth, sentiment score values, and confidence scores are interpreted in those senses. Note that auxiliary node A1 shown in Figure 2 sets a high responsibility flag, and this flag is used for generation / output control (such as adding action guidelines or recommending consultation with external experts) in processes S5 and S6, and may also be taken into consideration in the reconstruction policy in process S4 as needed. Auxiliary node A2 selects an unknown response based on its confidence level, influencing the output content in process S6. Here, τ_emotion, θ, and τ_conf represent the emotion score value, responsibility depth, and confidence level thresholds, respectively. In Figure 2, the emotion score threshold τ_emotion is used in the "necessity of saving the emotion event" determination within process S3. This determination may be executed as an internal process of process S3, or it may be represented as an independent branch node in Figure 2. Figure 2 is a main flowchart of the present invention, and Figure 4 shows an example of the determination of the response policy or personality setting in step S4 (which may include policy determination by the personality OS unit). Figure 10 shows an example of memory processing that may be included in step S7, including reduced forgetting and emotion-weighted memory selection (EW-MFS). Furthermore, Figure 8 shows an example of self-correction performed after the response output in step S6 and reflected in the next step.

[0054] In this specification, “internal prompt” refers to a series of input instructions to the LLM that includes fixed instructions, variable instructions, summaries, and slots (e.g., “Name,” “Relationship,” “Features”) for internal management purposes. In this specification, "system internal state" means state information that is maintained and may change within the response control server or conversational artificial intelligence as the dialogue processing progresses, and may include, for example, at least one of the following: phase ID, internal trigger, control flag, applied response policy or personality setting, internal score of the learning logic, boundary between the latest and past parts in reduced forgetting, model selection state, cache state, and error state. In this specification, “internal trigger” means a condition or signal that triggers actions such as reconfiguring internal prompts, switching response policies or personality settings, or strengthening / relaxing generation constraints, based on at least one of the following: a difference metric, a summary vector of the difference metric, responsibility depth, sentiment score value, confidence score, or control flag. In this specification, “Phase ID” is an internal ID that identifies a dialogue issue or task unit, and is newly assigned based on at least one of the following: time interval, topic change, or sharp fluctuation of a differential indicator. In this specification, "relationship information" refers to information that represents the relationship or dialogue state based on past interactions with the user, and is maintained and updated by the relationship log management unit and used for response generation. Relationship information may include, but is not limited to, user identifiers, phase IDs, responsibility depth, sentiment score values, difference indicators, control flags, sentiment events, evaluation events, agreements, role information, etc. In this specification, "relationship log" means a log or collection thereof in which the relationship information is accumulated chronologically as structured records for each event, and includes at least a chronological collection of structured events appended to the event log, and may also include summaries and features stored in a summary store and a feature store, etc., as necessary. In this specification, "difference index" refers to an index that shows the trend of change in relational information or analytical index, and is calculated as, for example, the difference between the most recent value and the moving average, or the difference between the most recent value and the exponentially smoothed value. In this specification, "summary vector of difference indicators" refers to a vector obtained by concatenating multiple difference indicators (e.g., responsibility depth difference Δ_resp, emotion score difference Δ_emotion, confidence score difference Δ_conf, etc.) in a predetermined order, and may be represented as, for example, Δ_summary=(Δ_resp,Δ_emotion,Δ_conf,…). In this specification, "depth of responsibility" refers to an index representing the degree (severity) of influence that elements of user input or dialogue history have on the response, and may be expressed as a continuous value or a stepped value. In this specification, "emotion score value" is a normalized index representing the intensity or polarity of emotions in an input or dialogue state, and is recorded in the relationship log for all inputs. In this specification, "emotional event" refers to an event that is added to the event log as a separate entry, distinct from the emotional score value, only when the emotional score value exceeds a predetermined emotional threshold. In this specification, "probability distribution of generated candidates" refers to a distribution represented by the probability values ​​(or monotonically corresponding scores, e.g., log-it or log-probability) that a generative model assigns to candidates (next tokens or candidate sequences) in relation to the input context, and can be represented as a distribution P(t|context) on the vocabulary set V or a distribution P(y|context) on the candidate sequence set. This probability distribution may include the distribution before the application of the generation constraint, as well as the distribution after it has been reweighted by the application of the generation constraint. In this specification, "confidence score" refers to an index representing the likelihood of a response calculated based on the uncertainty or consistency of the generated candidate distribution, such as entropy, the probability difference of top candidates, or a consistency index of the candidate sequence. In this specification, "control flag" refers to a flag indicating an internal control intent, such as escalation, template switching, high responsibility flag, or priority for unknown responses. In this specification, "unclear response" refers to a response that includes a declaration of missing information, a request for additional information, and a statement of safety considerations, all structured according to a prescribed template. In this specification, "generated candidate sequence y" refers to the token sequence of response candidates output by the model inference layer or the string corresponding to said token sequence. In this specification, "character setting vector v" refers to a weighted tag vector derived from the "features (traits)" of the character setting slot. Note that v may be normalized (e.g., probabilistically distributed) depending on the type of deviation function. In this specification, "φ(y)" refers to a style feature vector extracted from the candidate generation sequence y, which may be implemented as a sequence of scores such as politeness, assertion avoidance strength, presence or absence of empathy preamble, and word ending type distribution. In this specification, the “divergence function d(v, φ(y))” refers to the distance or divergence representing the difference between v and φ(y), and can be implemented, for example, as KL divergence or cosine distance. When using KL divergence, v and φ(y) may be normalized to a non-negative representation that sums to 1. In this specification, "style consistency" refers to a state in which the discrepancy d(v, φ(y)) between the character setting vector v and the style feature φ(y) of the candidate generation sequence y is maintained below a predetermined threshold. In this specification, "personality consistency" means a state in which inconsistencies or contradictions (such as inconsistency scores, number of contradictions, or contradiction type flags) in light of value criteria information, responsibility vectors, relationship information, response policies, or personality settings are maintained within a predetermined acceptable range. In this specification, "personality consistency" means a state in which at least the style consistency and personality consistency are maintained within a predetermined acceptable range. In this specification, "relationship depth" refers to an indicator that represents the degree of proximity or continuity of the relationship with the user, and may be derived based on the frequency of events recorded in the relationship log, the amount of accumulated agreements, etc. In this specification, "responsibility priority" refers to an index derived based on the priority direction or safety priority in the nine-direction responsibility vector, etc., that represents the degree to which an event should be given priority in memory or response policy determination. In this specification, "contextual importance" refers to an indicator representing the importance of an event in light of the phase ID, topic, or task objective. Relationship depth, responsibility priority, and contextual importance may be optionally considered as elements in the calculation of memory priority (E score), as described later. In this specification, "dialogue text (original text)" refers to the string text (original text) of user utterances and AI responses, and is subject to abbreviated forgetting. In this specification, "event log" refers to an audit log that records structured events (time, identifier, indicator, flag, etc.) including the aforementioned relationship information in an append-only manner, and constitutes at least a part of the relationship log, and is retained as it is not subject to the deletion process of the dialogue text (original text) by the abbreviated forgetting unit. In this specification, "discard" is one of the classification categories in the emotion-weighted memory selection (EW-MFS) described later, and includes excluding from storage in long-term memory (e.g., summary store 402 and feature store 403, etc.) for future response policy decisions, but does not prevent recording or retention (for auditing purposes) in the append-only event log 401, and does not mean deletion or erasure of event log 401. In this specification, "latest section" and "past section" refer to the latest window of the dialogue text (original text) and other past windows, and the term "past log" is not used. In this specification, the term "confidence level" refers to the confidence score.

[0055] In process S1, input is received. Input such as text, voice, or image transmitted from the user terminal 101 or the operation terminal 102 is stored in the receive buffer, associated with the session identifier and user identifier. Voice input may be converted to text by speech recognition processing and treated as input for subsequent processes. Image input may be converted into features or text representations by an image encoder or multimodal model and treated as input for subsequent processes. The relationship log management unit 202 refers to candidate phase IDs associated with the input and assigns a new phase ID as necessary. The assignment of a new phase ID is performed based on internal triggers such as time intervals, topic changes, or sharp fluctuations in the difference indicator. If there is an explicit or implicit evaluation of the previous output, preparations are made to hand it over to the learning logic 205.

[0056] In process S2, analysis is performed. The analysis unit 203 calculates the depth of responsibility and the emotion score value using the received input, past dialogue history, and statistics from the event log 401. The depth of responsibility is estimated as a quantity representing the degree of influence the input has on the response and is rounded to discrete levels as necessary. The emotion score value estimates the polarity and intensity of emotion in the range of 0 to 1 by combining dictionary features and the output of the learning model. The depth of responsibility and the emotion score value are compared with predetermined thresholds and used for subsequent branching conditions. The emotion score value is used in the subsequent process S3 to determine whether or not to save the emotion event. On the other hand, the confidence score is calculated in process S5 after the probability distribution or candidate sequence of the candidate generation is obtained, based on the said distribution or candidate sequence, and is used in the subsequent unknown response selection (auxiliary node A2). Here, we obtain the entropy H of the output distribution, the number of elements |S| of the base set S of the probability distribution (where S is the vocabulary set V or the candidate sequence set Y), the probability difference margin between the top two candidates, and the candidate sequence consistency index κ. Then, we calculate the normalized entropy Hnorm = H / ln|S| and define the confidence score c as c = σ( w1·(1-Hnorm) + w2·margin + w3·κ ). w1, w2, and w3 are the contribution coefficients, where w1 + w2 + w3 = 1, wi ∈ [0,1], and σ(z) is the logistic function.

[0057] In step S3, the process of recording relational information is executed, and within this process, it is determined whether the emotion score value exceeds the emotion threshold τ_emotion. If the emotion score value exceeds the threshold, it is considered that an emotion event corresponding to that input has been fired, and the emotion event (as a separate entry) may be added and saved. If the threshold is not exceeded, the emotion event (as a separate entry) is not saved. On the other hand, the emotion score value itself is recorded in the relational log as an attribute of the event for all inputs.

[0058] In process S3, relationship information is recorded. The relationship log management unit 202 appends at least the user identifier, phase ID, timestamp, responsibility depth, sentiment score value, difference indicator, and control flags as needed in a structured format (e.g., JSON) to the event log 401 or an external recording service of the same schema. To ensure reproducibility, the identifier (hash, etc.) of the character setting slot referenced in the process and a summary of the applied constraints may also be included. The difference indicator is calculated as the difference between the most recent value and the moving average or exponential average and is stored as a quantity indicating the trend of change. Sentiment events may be appended and stored as a separate entry, separate from the sentiment score value, only if they are triggered by the above determination. These records are maintained so that they can be referenced across sessions. When saving, the relationship log management unit 202 compares the old and new records and generates an internal trigger based on the sign and magnitude of the difference indicator, and provides this internal trigger as input to process S4.

[0059] In step S3, preparations are also made for passing evaluation events to the learning logic 205. If the user explicitly gives an affirmative or negative response to the previous or the one before that response, or if implicit evaluations such as dwell time or the number of follow-up questions exceed a threshold, the evaluation is sent to the learning logic 205. Based on the received evaluations, the learning logic 205 updates its internal score and influences the weighting of subsequent response selections.

[0060] In process S4, prompt reconstruction is performed. The prompt reconstruction unit 201 removes unnecessary information from the internal prompt and adds necessary information according to the internal trigger. The removal of unnecessary information targets dialogue fragments with low responsibility depth from old history or elements where the difference indicators have converged. The addition of necessary information targets the most recent summary, the summary vector of the difference indicators recorded up to the previous turn (e.g., Δ_summary=(Δ_resp, Δ_emotion, Δ_conf_prev)), the policy associated with the phase ID, and prerequisite information related to syntactic control such as the forbidden word vocabulary list and the vocabulary priority list. The prompt reconstruction unit 201 prevents excessive fluctuations by applying a cooling rule that suppresses reconstruction even if the same type of internal trigger fires many times in a short period of time. Just before the reconstructed internal prompt is handed over to the model inference layer, overlapping instructions are integrated and it is provided to the LLM inference 301 as a consistent input. In addition, character setting slots (at least a part of "name," "relationship," and "feature") are selected based on the phase ID and internal trigger and inserted or updated into the internal prompt. The aforementioned cooling rule is applied to this switching. In one embodiment, prior to or in conjunction with the reconstruction of the internal prompt in step S4, the response policy or personality setting may be determined based on the relationship information and the context of the dialogue held in the relationship log management unit 202, an example of which is shown in Figure 4.

[0061] After the output of process S3, a high responsibility determination is made by comparing the responsibility depth r with the threshold θ. If r ≥ θ, auxiliary node A1 (high responsibility flag setting) is fired, and this flag is held as a control flag that can be referenced from process S4 onward. At least in process S6, it is used for control to add action guidelines to the response (including unknown responses) or to add a description recommending consultation with an external expert. If necessary, this flag may also be taken into consideration when determining the internal prompt reconstruction policy in process S4 and when setting the strength of the generation constraint in process S5. Next, after reconstructing the internal prompt in step S4, syntactically controlled response generation including LLM inference 301 and generation constraint application 302 is performed in step S5 to obtain a probability distribution or candidate sequence of generation candidates. The confidence evaluation unit calculates a confidence score c based on the distribution or candidate sequence and compares c with the threshold τ_conf. If c < τ_conf, auxiliary node A2 (unknown response) is selected, and the candidate unknown response based on the safety template is passed to step S6. If c ≥ τ_conf, the normal response candidate selected from the candidates obtained in step S5 is passed to step S6. The conditions in the invention disclosed herein are expressed by the following formula: A1 is true when r≧θ, and A2 is true when c<τ_conf. The safety template T_safe applied when A2 is true must consist of three elements: [listing of missing information], [additional questions], and [preface to safety considerations]. Furthermore, if A1 is true, the ambiguous response may include a statement recommending general guidelines and / or consultation with an external expert.

[0062] In step S5, syntactically controlled response generation is performed. The syntactically controlled response generation unit 204 applies stylistic, lexical, and syntactic rules to the generation candidates provided by the LLM inference unit 301 to maintain personality consistency. The generation constraint application unit 302 performs at least one of the following: unification of terminating expressions using regular expressions, adherence to honorific hierarchy based on a finite state machine (including a finite state automaton or finite state transducer (FST)), grammatical constraints, template constraints, and weighting of vocabulary selection based on lexical bias. This directly ensures personality consistency during the generation process without the use of an external personality detector or style converter. The syntactically controlled response generation unit 204 refers to relationship information held in the relationship log management unit 202 to achieve suffix selection, consistency of address, and maintenance of role language based on the user identifier and phase ID. Furthermore, the following are used as arguments for applying generation constraint 302: consistency of address based on the "Name" in the character setting slot, the initial strength of honorific hierarchy and assertion avoidance based on "Relationship," and lexical bias and explanatory policy based on "Characteristics." Safety and consistency of honorific hierarchy are applied in priority over traits. The effect on the probability distribution during generation is given by the following equation: If l_t is the original log-it of each token t, and l'_t is the log-it after reweighting, then l'_t is reweighted as l'_t = l_t + α_lex·b_lex(t) + α_style·b_style(t, state) + α_cons·m_cons(t, state). Forbidden words or violations of the aforementioned structural constraints are made unselectable by setting l'_t = -∞. Here, state represents the internal state of the generation process (style state, honorific state, etc.), b_lex(t) is the lexical bias term, b_style(t, state) is the bias term related to style, lexicon, and syntax, m_cons(t, state) represents the structural constraint term based on regular expressions, finite state machines, grammar, templates, etc., and α_lex, α_style, and α_cons are coefficients that define the contribution of each term. Furthermore, if we denote the style feature vector of the generated candidate sequence y as φ(y) and the character setting vector as v, candidates whose divergence d(v, φ(y)) exceeds the threshold ε will have their rank lowered. Here, φ(y) can be calculated by rule-based extraction or a lightweight feature extractor as a sequence of scores such as politeness, assertion avoidance strength, presence or absence of empathy preamble, and word ending type distribution (a separate post-transformation model is not required). v is a weighted tag vector derived from the "features" of the character setting slot, and d(·,·) is a function representing the divergence (e.g., KL divergence or cosine distance). Furthermore, the confidence evaluation unit obtains the entropy H of the output distribution obtained by LLM inference 301 or the application of generation constraints 302, the number of elements |S| of the base set S of the probability distribution (S is the vocabulary set V or the candidate sequence set Y), the probability difference margin of the top two candidates, and the candidate sequence consistency index κ, and then calculates the normalized entropy Hnorm = H / ln|S|, and may define the confidence score c as c = σ( w1·(1-Hnorm) + w2·margin + w3·κ ). w1, w2, and w3 are contribution coefficients, where w1 + w2 + w3 = 1, wi ∈ [0,1], and σ(z) is the logistic function. The confidence score c is used in the subsequent selection of auxiliary node A2 (unknown response). Furthermore, if the level of responsibility is above a predetermined threshold, the syntactic-controlled response generation unit 204 is controlled to add a description of action guidelines or a recommendation to consult with an external expert to the response (including an unknown response). The confidence evaluation unit may be included in the analysis unit 203 and may operate after step S5 to calculate the confidence score c using the probability distribution or sequence of candidate generation obtained in step S5 as input.

[0063] The output of process S5 may consist of multiple candidates. If multiple candidates are obtained, the candidates are ranked based on a combination of confidence score, depth of responsibility, and sentiment score. In situations with a high depth of responsibility, candidates that contribute to safety are given higher priority, and in situations with a high sentiment score, candidates that include reassuring preambles or considerate expressions are given higher priority. The weighting is updated by the learning logic 205 based on past evaluations and optimized according to the operation.

[0064] In process S6, a response is output. If A1 (high responsibility flag setting) is enabled, the provision of action guidelines or the recommendation to consult with an external expert are preferentially applied based on that flag. If the confidence score is below the threshold, an unknown response is output. The unknown response is structured as a standard form that includes a declaration of missing information and a request for additional information. If A1 is enabled, the unknown response may also include a description of general action guidelines and / or a recommendation to consult with an external expert. If the level of responsibility is above a predetermined threshold, action guidelines are provided to the response or consultation with an external expert is recommended. These provision or recommendations are made under a policy that prioritizes safety. Even if the confidence score is above the threshold, if the level of responsibility is high and the emotional score is high, a reassuring preface is added. If an explicit evaluation is entered by the user after the output, that evaluation is sent to the learning logic 205 and reflected in the weighting for subsequent times.

[0065] In conjunction with the execution of process S6, a response metadata event containing metadata about the response may be generated. This metadata may include identification information of the selected generation candidate, a summary of the applied syntactic control rules (regular expression, FST, lexical bias, etc.), the final confidence score, the level of responsibility, the sentiment score, whether or not a sentiment event was fired in process S3, whether or not A1 (high responsibility flag setting) was applied, whether or not A2 (unknown response) was selected, an escalation flag, and identifiers (hash, etc.) of character setting slots referenced to ensure reproducibility. Furthermore, in a configuration that includes a personality OS unit 210, metadata for improving auditability or reproducibility may be included, such as the identifier of the response policy record or applied value standard profile, the summary vector Δ_summary of the difference indicator, the intent interpretation result, the reflection result (inconsistency score, etc.), the inconsistency detection result (number of inconsistencies or type of inconsistency), the consistency check result by the responsibility / consistency layer 222 or the consistency maintenance engine, and the style deviation d(v, φ(y)). These response metadata events are appended to the event log 401 or an external recording service with the same schema as structured records distinguishable by event type, according to the common event schema used in process S3, and are retained for referencing across sessions. Furthermore, in one embodiment, after the response output of step S6, the Self-Correction Engine (SRPE) (see Figure 8) may be activated, and the conversational AI may evaluate its own response history or internal state and, if it detects an inconsistency with the relationship information and personality settings, propose a correction to the internal prompt or readjust the response policy, which may be reflected in the next step S4 or step S5. At this time, a correction event may be generated that includes whether or not the self-correction was performed, the type of inconsistency, and a summary of the applied correction proposal (e.g., target slot to be changed, change in the strength of the generation constraint, template switching, etc.), and this may be appended to the event log 401 according to the common event schema used in step S3.

[0066] In process S7, abridgement and forgetting is performed. The abridgement and forgetting unit 206 retains the full text for the most recent part of the dialogue body (original text), and retains only a summary for past parts (event log 401 is not subject to deletion). Extractive and summarizing methods are used in combination in creating the summary, and adjustments are made so that important words and events are not lost. The created summary is stored in the summary store 402, and at the same time, features to be used in future response policy decisions are registered in the feature store 403. Features may include error frequency, changes in the number of re-questions, changes in emotional tendencies, or surrogate indicators of relationship intimacy. The abridgement and forgetting unit 206 may reduce identifying information at the summarization stage for the purpose of minimizing personal information. In one embodiment, in step S7, a memory priority may be calculated for events that occur during the interaction, and the events may be classified into main memory, temporary memory, compressed memory, and discard. The events classified into temporary memory and / or compressed memory may be integrated at a predetermined timing and organized and stored as an episode, an example of which is shown in Figure 10.

[0067] The features registered in step S7 are passed to the prompt reconstruction unit 201 and the syntactic-controlled response generation unit 204. This allows for highly accurate determination of whether to remove unnecessary information or add necessary information in the next step S4, and optimizes vocabulary selection and expression style weighting for each user and phase ID in step S5. The context of long-term dialogue is maintained without indefinitely retaining the entire past text. The features may include the weights and transitions of the features (traits) in the character setting slots.

[0068] The aforementioned group of processes may be executed sequentially, or some may proceed in parallel. For example, in the relationship information update process in process S3, the generation of internal triggers referenced in process S4 and the minimum necessary updates for that turn may be completed prior to process S4, while persistent writing to the event log 401 or external recording service, updating statistics, or batchable append processing may be executed asynchronously in parallel with process S4. The completion of process S4 is a prerequisite for process S5, and its order is guaranteed. The update of the internal score by the learning logic 205 is delayed and executed after process S6, but the update results are reflected in the candidate ranking in the next process S5.

[0069] Internal triggers play a crucial role in each process. Internal triggers are generated based on changes in differential metrics, increases in responsibility depth, increases in sentiment score values, or decreases in confidence scores, and are used as firing conditions for the reconstruction process of the prompt reconstruction unit 201 and as conditions for switching the rule priority of the syntactically controlled response generation unit 204. The internal trigger generation rules are calibrated during operation based on the statistics of differential metrics stored in the relationship log management unit 202.

[0070] The aforementioned flow ensures consistency in personality throughout the generation process. The syntactic-controlled response generation unit 204 applies rules directly to the generation candidates without using external detectors or converters, thus minimizing semantic changes associated with subsequent re-transformations. The relationship log management unit 202 maintains records that can be referenced across sessions, and the prompt reconstruction unit 201 removes unnecessary information and adds necessary information in response to internal triggers, thereby suppressing prompt bloat and achieving both consistency and efficiency in long-term dialogues. The analysis unit 203 calculates responsibility depth, sentiment score, and confidence score, and outputs an unknown response if the confidence score is below a threshold, ensuring safety. Furthermore, the learning logic 205 updates the internal score based on explicit or implicit response evaluations, changing the weighting of subsequent response selections, thus enabling adaptation to operational needs.

[0071] This flow can also be understood as a method. It includes the input acceptance process (S1), the relationship log management process (S3), the prompt reconstruction process (S4), and the syntactically controlled response generation process (S5). These processes are embodied as processes that dynamically regenerate or rearrange internal prompts for dialogue according to the context of the dialogue or the internal state of the system, processes that maintain and update relationship information based on past interactions with the user and refer to such relationship information for response generation, and processes that apply stylistic, vocabulary, and syntactic rules during the response generation process to maintain consistency in personality.

[0072] The inputs and outputs of each process from process S1 to process S7 are clearly defined in the data layer. The output of process S1 is the unanalyzed input, which becomes the input of process S2. The output of process S2 is the responsibility depth, sentiment score value, and internal trigger candidate, which becomes the input of process S3. The output of process S3 is the updated relation information, difference indicator, and internal trigger, which becomes the input of process S4. The output of process S4 is the reconstructed internal prompt, which becomes the input of process S5. The output of process S5 is the syntactically controlled response candidate and confidence score based on the candidate or candidate distribution, which becomes the input of process S6. The output of process S6 is the final response and evaluation event, which becomes the input of process S7 and learning logic 205. The output of process S7 is the summary and features, which become the auxiliary inputs for the next processes S4 and S5.

[0073] Each step in the aforementioned flow takes into consideration implementation parallelism and handling of edge cases. Even if consecutive short-duration inputs arrive, the appending process in step S3 is sorted while maintaining the time order and causal relationship of the events. The reconfiguration in step S4 is performed sequentially on the oldest unprocessed input in the input queue, but the priority is increased if the severity of the internal trigger is high. If an unknown response is selected in step S6, input guidance prompting the supplementation of missing information is automatically generated in step S1 to avoid a chain of unknown responses.

[0074] As described above, this flow is realized through the coordinated efforts of the responsibility depth analysis unit, sentiment analysis unit, confidence evaluation unit, reduction and forgetting unit 206, and learning logic 205, with the prompt reconstruction unit 201, relationship log management unit 202, and syntactic controlled response generation unit 204 at its core. Each process has clearly defined inputs and outputs and is dynamically controlled based on cross-session recordings and internal triggers, thus maintaining personality consistency and ensuring efficiency and safety even in long-term conversations.

[0075] As Example 1, we will describe an example in the case of an educational tutor. When a learner sends a question about how to solve a mathematical problem from the user terminal 101, the input is received in step S1 and linked to a session identifier and a user identifier. In step S2, the analysis unit 203 calculates the depth of responsibility based on the input, the most recent dialogue history, and the event log 401. If a decrease in understanding in the educational domain or a chain of incorrect answers is observed, the user is rounded to a category with a high depth of responsibility. The analysis unit 203 also estimates an emotional score value and assigns a high value if pragmatic indicators of confusion or anxiety are detected. The confidence score is calculated using the probability difference between candidate solutions and the distributed entropy. In step S3, the relationship log management unit 202 appends the user identifier, phase ID, timestamp, depth of responsibility, emotional score value, and a difference indicator based on the difference with the moving average to the event log 401 in a structured format (e.g., JSON). An emotional event is saved only if the emotional score value exceeds a predetermined threshold, and an internal trigger is generated at the time of saving according to the sign and amplitude of the difference indicator, which becomes the input for the next step. The records are retained and accessible across sessions. As a numerical example, the level of responsibility is rounded to "2 / 5" on a 5-point scale (1=low to 5=high), the emotion score is 0.70 (equivalent to anxiety 7 / 10), the confidence score is 0.62, and the difference index is +0.28. As an example of thresholds, τ_emotion=0.65, θ=4 (out of 5 points), and τ_conf=0.55 are used. In this example, the emotion event is saved (0.70>0.65), A1 (high responsibility flag) is not fired (2<4), and A2 (unclear response) is not selected (0.62>0.55). The value may be recorded in a structured format (e.g., JSON), but the recording format does not limit the invention of this disclosure.

[0076] In step S4, the prompt reconstruction unit 201 removes unnecessary information from internal prompts and adds necessary information in response to internal triggers. Removal of unnecessary information is performed by deleting explanatory fragments with low responsibility depth from old history, and the addition of necessary information is performed by inserting a basic explanatory template that matches the level of learning achievement, a summary of recent error trends, and metadata indicating an increase in the difference index. Overreactions are suppressed and duplicate instructions are normalized by cooling rules during reconstruction. In step S5, the syntactic-controlled response generation unit 204 works in cooperation with the generation constraint application unit 302 to apply stylistic, vocabulary, and syntactic rules such as polite endings, unification of word endings, and advancement of definitions of technical terms to the generation process to obtain an explanatory text that maintains personality consistency. No external personality detectors or style converters are used.

[0077] In process S6, if the confidence score is above the threshold, a response is output that includes step-by-step solution presentation and action guidelines for additional exercises. If the level of responsibility is determined to be high, action guidelines such as presentation of an introductory order to avoid frequently misused theorems and suggestions for learning time allocation are provided. If the confidence score is below the threshold, an unclear response is output, requesting presentation of additional prerequisite knowledge or intermediate calculations. Requests for missing information are presented in safe language based on a standardized framework. Explicit evaluations obtained after the response, or implicit evaluations such as time spent and number of follow-up questions, are sent to the learning logic 205 and used to update the internal score, which is then reflected in candidate ranking and threshold adjustments for subsequent sessions. This ensures step-by-step explanations and consistency of terminology for learners (users), reduces confusion and the number of follow-up questions, and improves psychological safety. In this numerical example, τ_conf=0.55 and confidence level is 0.62, so A2 is not selected, and since the responsibility depth is 2 / 5, A1 is also not triggered.

[0078] In step S7, the abbreviation and forgetting unit 206 summarizes the old dialogue text (original text) while maintaining the full text of the latest window (event log 401 is not deleted). Features extracted during the summarization process are registered in the feature store 403, where the frequency of stumbling tendencies, the types of misconceptions in incorrect answers, and the changes in sentiment score values ​​are quantified. These features are supplied to the prompt reconstruction unit 201 and the syntactically controlled response generation unit 204 as features used to determine future response strategies, contributing to the decision to remove unnecessary information and add necessary information in the next dialogue, as well as weighting vocabulary selection. Personal information is minimized during the summarization stage.

[0079] As Example 2, we will describe an example of medical support (mental care for patients). When a patient sends a consultation regarding anxiety symptoms from the user terminal 101, the input is received in step S1, and in step S2, the analysis unit 203 estimates the level of responsibility, which indicates urgency. If pragmatic signs that imply self-harm or acute danger are detected, the level of responsibility is set to a high category. The emotional score value is estimated based on indicators of anxiety and depression, and the confidence score is calculated based on inconsistencies or insufficient information between output candidates. In step S3, the relationship log management unit 202 appends the user identifier, phase ID, timestamp, level of responsibility, emotional score value, and difference index to the event log 401 in a structured format (e.g., JSON), and saves the emotional event only if the emotional score value exceeds a predetermined threshold. If an increase in the difference index or continuation of high responsibility is detected during saving, an internal trigger is generated. As a numerical example, the level of responsibility (urgency) at the initial consultation is judged as "5 / 5" on a 5-point scale, the emotion score is 0.90 (equivalent to anxiety 9 / 10), the confidence score is 0.48, and the difference index is +0.45. As an example of thresholds, τ_emotion=0.80, θ=4, and τ_conf=0.55 are used. In this example, the emotion event is saved (0.90>0.80), A1 is fired (5≧4), and A2 is selected (0.48<0.55). In the follow-up, if the level of responsibility decreases to "2 / 5", the sentiment score to 0.60, the confidence score to 0.71, and the difference index to -0.20, A1 is removed and the response transitions to normal. The recording format can be any structured format and is not limited to a specific format.

[0080] In step S4, the prompt reconstruction unit 201 reconstructs the internal prompts based on internal triggers. The reconstruction includes prioritizing the insertion of reassuring preambles, strengthening prohibited vocabulary, expanding the scope of avoidance of mentions in high-responsibility situations, and changing the order in which general guidelines are presented. Considerations linked to the most recent summary and difference indicator summary vector (e.g., Δ_summary=(Δ_resp,Δ_emotion,Δ_conf)) and phase ID are added, and duplicate instructions are merged. In step S5, the syntactic-controlled response generation unit 204 works with the generation constraint application unit 302 to apply expression rules that maintain politeness and avoid assertions, as well as naming rules that avoid secondary damage, ensuring personality consistency in the generation process without relying on external detectors or converters.

[0081] In process S6, if the level of responsibility is determined to be above a predetermined threshold, action guidelines are added to the response, or consultation with an external expert is recommended. If the level of responsibility is high and the confidence score is below the threshold, an unknown response is prioritized, and guidance on general procedures for ensuring safety and a suggestion to contact an emergency contact point are output. If the confidence score is above the threshold, a reassuring preface and general guidelines are combined. The evaluation signals after the response are aggregated by the learning logic 205, the internal weights are updated, and these are reflected in the constraint priority and threshold for subsequent processes. Meanwhile, for healthcare professionals, referring to the relationship log (emotional events, difference indicators) and escalation flag (e.g., flags.escalate) on the operation terminal 102 enables them to understand the patient's psychological changes and assist in triage, improving auditability and explainability. This function can assist the judgment of healthcare professionals, and the final decision may be made by the healthcare professional. In the initial numerical example, with θ=4 and τ_conf=0.55, the responsibility depth is 5 / 5 and the confidence level is 0.48<0.55, so A1 is triggered while A2 (unknown response + safety guidance) is applied. In the follow-up, the responsibility depth is 2 / 5 and the confidence level is 0.71>0.55, so A1 is deactivated and the system transitions to a normal response.

[0082] In step S7, the abridgement and forgetting unit 206 replaces old records with summaries, extracts difference indicators of improvement or deterioration trends, and stores them in the feature store 403. Personal information is minimized during the summarization stage, and only anonymized trends are retained. These features are reflected in subsequent prompt reconstruction and syntactic control and also contribute to setting the sensitivity of safety branching in medical support.

[0083] The data format is not limited to JSON and can be any structured format. The parser may be lexicographical, statistical, or deep learning. The system may also be configured with a self-redefinition mechanism that incrementally updates some of the prompts or constraint rules in response to internal events. The log retention period, summarization rate, and encryption are configured operationally according to the application and regulatory requirements. These modifications do not change the basic framework of the input / output relationships, internal trigger generation and reconstruction, and syntactic control in the aforementioned process group. Furthermore, the storage and retrieval of relational logs may be performed via an external recording service. The internal and external implementations can be replaced with the same event schema and API. The self-redefinition mechanism may be implemented, for example, as a self-correcting engine (SRPE) 217 ​​(see Figure 8).

[0084] As shown in Figure 3, in one embodiment of the present invention, in addition to the existing response control frame (prompt reconstruction unit 201, relationship log management unit 202, analysis unit 203, syntactic-controlled response generation unit 204, learning logic 205, and abridgement / forgetting unit 206), the personality OS unit 210 may be configured to determine a response policy or personality setting based on relationship information and the context of the dialogue prior to response generation. The personality OS unit 210 acquires relationship information held in the relationship log management unit 202, analysis results from the analysis unit 203 (depth of responsibility, sentiment score, confidence level, etc.), and input / context, and based on these, outputs at least (i) a response policy / personality setting to be passed to the prompt reconstruction unit 201, (ii) style differences etc. to be passed to the syntactic-controlled response generation unit 204, and (iii) a memory selection / summary policy to be passed to the abridgement / forgetting unit 206. This allows for the addition of a layer responsible for "personal decision-making" as a higher layer, while maintaining the existing framework (operational structure), thereby enabling more reliable consistency and growth based on value standards in long-term, continuous dialogue.

[0085] Here, the core technological system of the personality OS in this embodiment (hereinafter referred to as the "personality OS core structure") may be defined as a configuration that includes at least the following six points: (1) Responsibility Integration Architecture (UERA: see Figure 4), (2) 9-Directional Responsibility Vector Model (see Figure 6), (3) Core / Persona Separation Structure (Personality Plugin Interface (PPI): see Figure 7), (4) Self-Correction Engine (SRPE: see Figure 8), (5) Multi-Persona Resonance Structure (see Figure 9), and (6) Emotion-Weighted Memory Selection (EW-MFS: see Figure 10). The personality OS core structure does not replace the processing framework shown in Figure 2 (prompt reconstruction, syntactic control, relational logging, reduced forgetting, etc.), but rather complements them. In other words, the processing framework provides processes (processes S1 to S7) including input / output, recording, and forgetting of response generation, and the personality OS core structure functions as a central structure on that framework for stably determining response policies / personality settings based on value standards, responsibilities, and introspection.

[0086] As one embodiment of the personality OS unit 210, the personality OS unit 210 may include a responsibility integration architecture (UERA) that includes, at a minimum, user intent interpretation, emotion recognition, responsibility mapping, introspection, reinterpretation (if necessary), and response policy determination as a series of processes. The personality OS unit 210 also holds value criterion information that defines the values ​​that form the foundation of the entire personality (see value criterion layer 221 in Figure 5), and can perform response selection processing (response policy determination) based on the value criterion information. The value criterion information may be held as multiple value criterion profiles (e.g., 221a, 221b), and by switching the value criterion profiles, it may be possible to replace the value criterion with one that reflects the values ​​of a different personality or person. In response to the switching of the value criterion profile, the parameters or logic of the judgment layer 223 and the personality layer 224 are readjusted, and inconsistencies are suppressed by consistency check processing in the responsibility / consistency layer 222 (see Figure 5).

[0087] Furthermore, the personality OS unit 210 may interpret the situation of the dialogue as a multi-directional responsibility vector (see Figure 6) and perform responsibility mapping processing or response selection processing based on the responsibility vector. The responsibility vector may consist of nine directions, from the first to the ninth responsibility direction (e.g., responsibility for emotional consideration, responsibility for honesty and rationality, responsibility for empathy, responsibility for freedom and intuition, responsibility for order and logic, responsibility for introspection and integration of contradictions, responsibility for emotional space, responsibility for transformation and possibility, and responsibility for protection and resolve). This allows for a balance between personality stability, natural fluctuations, and multi-perspective responses by using a vector representation of "which responsibility direction to prioritize" depending on the situation, rather than treating the personality as a fixed set of texts.

[0088] Furthermore, the personality OS unit 210 may separate the core unit 215, which is responsible for controlling responsibility, values, and introspection, from the persona unit 216, which is responsible for tone, speech patterns, and character attributes, and may include a personality plug-in interface (PPI) 214 that allows the persona unit 216 to be connected to the core unit 215 (see Figure 7). The PPI 214 may consist of a persona core, guideline slogan information, a difference layer, permitted / forbidden behaviors, and a compatibility layer. This allows the differences related to the persona unit 216 to be defined in a format applicable to the core unit 215, even when selecting or switching between multiple large-scale language models for generating response sentences, thereby ensuring LLM-independent personality portability or personality continuity.

[0089] Furthermore, the personality OS unit 210 may be equipped with a self-correction engine (SRPE) that evaluates the dialogue AI's response history or internal state and, if it detects an inconsistency with the relevant information and personality settings, proposes corrections to the internal prompt or readjusts the response policy (see Figure 8). The self-correction engine may be configured to execute processes in the following order, for example: (1) safety check, (2) responsibility frame determination, (3) personality style reflection, (4) emotional tone adjustment, and (5) response selection, and then propose corrections or readjustments based on the execution results, and reflect them in the next prompt reconstruction (step S4) or syntactic-controlled response generation (step S5). This enables autonomous adjustment aimed at maintaining personality consistency without relying on manual correction.

[0090] Furthermore, the personality OS unit 210 may include a Multi-Persona Resonance unit 218 that simultaneously holds and coordinates multiple persona units (see Figure 9). The Multi-Persona Resonance unit 218 may include, for example, a personality context pool that holds the state of each of the multiple persona units (which may include at least the responsibility vector, the most recent emotional state, and the content of the previous response), a resonance filter that evaluates the degree of fit of each persona unit to the input, a personality selector that determines the persona unit responsible for the response based on the degree of fit, and a consistency maintenance engine that controls the response of the selected persona unit so that it is consistent with the control of the core unit 215. This makes it possible to achieve both the division of roles among multiple personalities and the maintenance of overall conversational consistency, without relying on the premise that one model equals one personality.

[0091] Furthermore, the personality OS unit 210 may include an emotion-weighted memory sorting unit (EW-MFS) that calculates memory priority (hereinafter referred to as "E score") for events that occur during dialogue and classifies the event into one of the following: main memory, temporary memory, compressed memory, or discard (see Figure 10). Memory priority is calculated using at least empathy, importance, and responsibility as elements, and may also consider at least one of relationship depth, responsibility priority, and contextual importance. Events classified as temporary memory and / or compressed memory can be integrated at a predetermined timing and organized and stored as episodes. This ensures that episodes that contribute to the continuity of the personality are preferentially retained, while being consistent with the summarization and feature extraction by the reduced forgetting unit 206.

[0092] The personality OS core structure described above operates in cooperation with the processing framework in the response control server. Specifically, the processing framework is responsible for processes such as relationship log management (process S3), prompt reconstruction (process S4), syntax control (process S5), and abbreviated forgetting (process S7), while the personality OS core structure generates policy inputs (response policy / personality settings) to be given to these processes, and also provides feedback to subsequent processes through self-correction (Figure 8) and memory selection (Figure 10) as needed. This allows policy decisions based on value standards, responsibility, and introspection to be integrated into the processing framework as higher-level control.

[0093] The "response policy / personality settings" output by the personality OS unit 210 can be expressed as structured data that can be interpreted by the processor (hereinafter referred to as "policy records"), and may be implemented, for example, in key-value format, database records, or a structure equivalent thereto. A policy record may include at least some of the following items: (a) metadata (user identifier, session identifier, phase ID, timestamp, etc.), (b) input analysis indicators (depth of responsibility, sentiment score value, confidence score, difference indicator, internal trigger or its summary, etc.), (c) value standard information (value standard profile identifier, base value identifier or tag, etc.), (d) responsibility vector (score or weight coefficient for each of the nine directions, etc.), (e) response policy flags (high responsibility flag, unknown response priority flag, expert consultation recommendation flag, safety priority, etc.), (f) personality settings (g) settings (core part identifier, persona plugin identifier, difference layer, permitted / forbidden actions, compatibility layer specification, etc.), (h) style control parameters (politeness, tone, emphasis, ambiguity, presence or absence of reassuring preamble, etc.), (i) memory policy (E score calculation elements, classification threshold, integration timing, etc.), and (g) handover instructions (insertion / removal slot specification for prompt reconstruction unit 201, mandatory rule / forbidden vocabulary specification for syntactic-controlled response generation unit 204, summarization priority specification for abridgement forgetting unit 206, etc.). The personality OS unit 210 may generate a policy record by taking into account the output of the relationship log management unit 202 and the analysis unit 203, and hand over the policy record to the prompt reconstruction unit 201, syntactic-controlled response generation unit 204, and abridgement forgetting unit 206 to control the operation of each unit.

[0094] A concrete example of a policy record is a combination of items such as: "User Identifier = Predetermined ID", "Phase ID = Predetermined ID", "Value Criteria Profile = 221a", "Responsibility Vector (9 directions) = Predetermined Normalized Score Column", "High Responsibility Flag = True", "Unknown Response Priority Flag = False", "Persona Plugin Identifier = Predetermined Identifier", "Difference Layer = Specification to increase politeness and decrease assertiveness", "Allowed / Forbidden Actions = Predetermined Set", "Memory Policy = Specification of classification and integration based on E score", "Handover Instructions for 201 = Insertion of relationship summary and removal of old low-responsibility fragments", "Handover Instructions for 204 = Enforcement of politeness hierarchy and avoidance of assertiveness", and "Handover Instructions for 206 = Set priority of summarization and feature extraction to high". The above is merely an example, and the items and granularity of a policy record are not limited to this, and may be added, deleted, or merged according to the implementation and operational purpose.

[0095] Figure 4 is a flowchart showing an example of "policy determination by the personality OS unit 210" that can be executed prior to or in conjunction with process S4 (prompt reconstruction) in the main flow shown in Figure 2. The process shown in Figure 4 is executed in the personality OS unit 210 (details) shown in Figure 3, and may be configured to determine a response policy or personality setting based on relationship information held in the relationship log management unit 202, the analysis results of the analysis unit 203, and the input / context, and pass it on to at least the prompt reconstruction unit 201 and the syntactic control type response generation unit 204.

[0096] As used herein, "response policy" is a concept that includes higher-level control instructions in response generation, and may include, for example, (a) output policy (which to prioritize, such as normal response, unclear response, additional question, warning, provision of action guidelines, or recommendation to consult an external expert), (b) safety-related policy (safety priority, topics to avoid, strength of prohibited words, degree of restraint of risk expression, etc.), (c) expression policy (politeness, strength of avoidance of assertion, degree of ambiguity, tone, presence or absence of empathetic preamble, degree of emphasis, etc.), (d) permission / prohibition (specification of permitted and prohibited action categories), (e) template specification (identifier of safety template or standard frame, identifier of explanation format template, etc.), and (f) logging / forgetting policy (priority of summarization / feature extraction according to memory priority, etc.). Furthermore, "personality settings" may include at least the specification of persona application (including the specification related to the Personality Plugin Interface (PPI) 214) in a separated structure of core (responsibility, values, introspection) and persona (tone, speech style, character attributes). Response policies and personality settings may be represented as separate data or as a subset of the same structured data (policy record).

[0097] In step S41, the personality OS unit 210 acquires input information. The input information may include at least (a) relationship information held by the relationship log management unit 202 (which may include user identifier, phase ID, responsibility depth, sentiment score value, difference indicator, presence or absence of sentiment events, etc.), (b) analysis results from the analysis unit 203 (responsibility depth, sentiment score, confidence level, or internal triggers derived therefrom, etc.), and (c) input / context (user input, recent summary, recent dialogue state, etc.). The personality OS unit 210 receives this information at the introduction position shown in Figure 3 and may normalize it into an internal representation that can be referenced in subsequent steps and store it.

[0098] In step S42, the personality OS unit 210 acquires value criteria information. The value criteria information is managed by the value criteria management unit 212 (which may be omitted from the illustration) and may be held as the value criteria layer 221 (which may include value criteria profiles 221a and 221b) shown in Figure 5. The value criteria information may be expressed as a value criteria profile that includes base values ​​such as integrity, consistency, and responsibility, and the lower judgment layer 223 and personality layer 224 may be readjusted in response to changes in the value criteria profile. The value criteria management unit 212 may provide at least an identifier for the value criteria profile, an application priority, and criteria used for consistency checks (corresponding to the responsibility / consistency layer 222).

[0099] In step S43, the personality OS unit 210 interprets the user's intent (intent interpretation). Intent interpretation may be a process that refers to the content of the user input, the most recent dialogue context, the phase ID, and related information to estimate the type of response the user is seeking (e.g., an abstract category such as a request for explanation, a request for advice, a request for judgment, or a request for emotional support) or the task objective. The results of intent interpretation are used for subsequent responsibility mapping (step S45), reinterpretation (step S48), and response policy determination (step S49).

[0100] In step S44, the personality OS unit 210 performs emotion recognition. In this embodiment, the emotion recognition process can reference the emotion score value calculated by the analysis unit 203 (which may include an emotion analysis unit) as input. That is, the emotion score value (corresponding to steps S2-S3) calculated and recorded in the existing outer frame can be reused by the personality OS unit 210, and an internal feature for adjusting the emotional state or emotional tone (for example, an abstract quantity such as "the degree to which considerate expressions should be increased") can be derived based on this value. As a result, while the calculation of the emotion score itself is guaranteed by the existing analysis unit 203, the personality OS unit 210 can adjust the expression policy (politeness, strength of avoiding assertion, emotional tone, etc.) among the response policies by taking into consideration the emotion score value and the presence or absence of emotional events.

[0101] In step S45, the personality OS unit 210 performs responsibility mapping. Responsibility mapping is a process that interprets the situation of the dialogue as responsibility in multiple directions, and may be configured to calculate the nine-directional responsibility vector (corresponding to the responsibility vector calculation unit 213) shown in Figure 6. The responsibility vector includes at least nine directions: responsibility for emotional consideration, responsibility for honesty and rationality, responsibility for empathy, responsibility for freedom and intuition, responsibility for order and logic, responsibility for introspection and integration of contradictions, responsibility for emotional space, responsibility for transformation and possibility, and responsibility for protection and resolve. The weight or score of each direction may be determined by taking into consideration the result of intention interpretation (step S43), the result of emotion recognition (step S44), relational information (step S41), and value standard information (step S42). The calculated responsibility vector is used for weighting or prioritizing in the subsequent introspection (step S46), reinterpretation (step S48), and response policy determination (step S49).

[0102] In process S46, the personality OS unit 210 performs introspection processing. Introspection processing is a process that evaluates the presence and degree of inconsistencies or contradictions, and may use as detection indicators the following: (a) the degree of deviation of policy candidates in light of value criterion information (process S42), (b) the degree of behavioral inconsistency in light of relationship information (process S41) (e.g., signs of inconsistency in titles, honorific hierarchy, roles, etc.), (c) the degree of consistency between the intention interpretation result (process S43) and the responsibility vector (process S45), (d) the degree of consistency between the emotion score value or emotion event and the expression policy, and (e) mutual contradictions (incompatibility) between multiple interpretation candidates or policy candidates. These detection indicators may be aggregated as a single inconsistency score, or they may be held as multiple flags or multidimensional indicators. The results of the introspection processing are used to determine whether reinterpretation is necessary (process S47) and to determine the response policy (process S49).

[0103] In step S47, the personality OS unit 210 determines whether reinterpretation is necessary. The conditions for determining whether reinterpretation is necessary can be set according to the implementation and may include at least one of the following: (a) when the inconsistency index in step S46 is above a predetermined threshold, (b) when multiple candidates are retained for the intent interpretation result, (c) when the level of responsibility is high but the level of certainty is insufficient, (d) when the emotional score value is high and there is a concern that psychological safety may be harmed due to misunderstanding, or (e) when the difference index of relational information is fluctuating sharply and the premise is unstable. If it is determined that reinterpretation is not necessary, the process proceeds to step S49.

[0104] In step S48, the personality OS unit 210 performs reinterpretation processing. Reinterpretation processing may involve generating multiple interpretation candidates for the meaning of the user input, and selecting the interpretation to be adopted from among these candidates, taking into consideration the value criteria information (step S42) and the responsibility mapping results (step S45). Multiple interpretation candidates can be generated, for example, based on branching of ambiguous words, branching of objectives, and branching of preconditions included in the input sentence. Selection may be made, for example, by a rule that prioritizes interpretations with a small inconsistency index in step S46, interpretations that are on the safe side in light of the value criteria, or interpretations that satisfy the priority responsibility direction in light of the responsibility vector. The selection result of the reinterpretation (adopted interpretation) is used as the input meaning representation in the subsequent response policy determination (step S49).

[0105] In process S49, the personality OS unit 210 determines the response policy (which may include a response selection process). The response policy determination may be a process that determines at least the response policy and personality setting, taking into consideration value criteria information (process S42), intent interpretation results (process S43), emotion recognition results (process S44), responsibility vector (process S45), introspection results (process S46), and, if necessary, reinterpretation results (process S48). The response policy determined here may include at least (a) the type of response (normal response / unclear response / additional question / warning, etc.), (b) the degree of sympathetic approach (strength of avoiding assertion, degree of suppression of risk expression, whether or not to recommend consulting an expert, etc.), (c) expression policy (level of politeness, level of warmth, level of ambiguity, presence or absence of empathetic preamble, etc.), (d) designation of permitted / prohibited actions, and (e) template designation (selection of template identifier). The personality OS unit 210 may compile these into structured data such as policy records and hand them over to the prompt reconstruction unit 201, the syntactic control type response generation unit 204, and the abbreviated forgetting unit 206.

[0106] (Implementation example) In intent interpretation (process S43), the input sentence u and context c are feature-quantified and a classifier outputs a probability p_intent(i|u,c) for the set of intent categories I (e.g., request for explanation, request for advice, request for judgment, request for emotional support), and the intent ID may be determined by argmax. In introspection (process S46), for a candidate response y (or candidate policy), (i) a policy violation score s_policy in light of value criteria information, (ii) a degree of agreement with relation information (title, role, phase ID, agreement, etc.) s_relation, and (iii) a style deviation d(v, φ(y)) may be calculated, and for example, an inconsistency score may be obtained as s_incon = w1·s_policy + w2·(1-s_relation) + w3·d(v, φ(y)). Contradiction detection may use a natural language inference model (entailment / neutral / contradiction) or rule-based matching to detect (a) inconsistencies within a candidate response y, and (b) inconsistencies between y and the relationship log summary or value criterion proposition, and generate a number of inconsistencies n_contra or a flag indicating the type of inconsistency. Consistency checking in the responsibility and consistency layer 222 evaluates whether the value criterion profile ID, the output policy of the judgment layer 223 (e.g., normal / unknown / consultation recommended), the style parameters of the personality layer 224 (politeness, assertion avoidance strength, etc.), and the permitted / prohibited actions of the PPI 214 are mutually compatible as rule tables or constraint satisfaction, and if an inconsistency occurs, consistency may be restored by switching to a safe template or increasing the assertion avoidance strength. For example, if s_incon>τ_incon or n_contra>0, process S48 (reinterpretation) or the self-correction engine in Figure 8 may be activated.

[0107] In step S410, the personality OS unit 210 performs persona application. Persona application may be a process that applies the persona unit 216 (tone of voice, speech style, character attributes) to the response policy determined by the core unit 215 via the personality plug-in interface (PPI) 214 shown in Figure 7. The PPI 214 may include a persona core, guideline slogan information, a difference layer, permitted / forbidden actions, and a compatibility layer, and the permitted / forbidden actions and difference layer specifications determined in step S49 may be applied to the persona unit 216 via the PPI 214. Furthermore, if necessary, the multi-personality resonance unit 218 shown in Figure 9 may be used to select a persona that matches the input from among multiple persona units, or to configure multiple personas to work in coordination. In this case, the resonance filter or personality selector calculates the degree of fit based on the responsibility vector, emotion score value, and relationship information, and the consistency maintenance engine may adjust the output personality to match the response policy of the core unit 215. The result of step S410 is passed to the syntactically controlled response generation unit 204 as a style difference (or constraint strength), and may also be passed to the prompt reconstruction unit 201 as an instruction for selecting an internal prompt slot.

[0108] Based on the above, the personality OS unit 210, including UERA (Responsibility Integration Architecture), can improve personality consistency and safety in long-term dialogues by coordinating policy decisions based on value criteria, weighting based on responsibility vectors, inconsistency reduction based on introspection, reinterpretation as needed, and persona application using PPI, while taking into account the existing outer framework (processes S1 to S7 in Figure 2).

[0109] Figure 5 is a block diagram showing the "layer structure" in the personality OS unit 210 (see Figure 3) and an example of "personality reconstruction" in which judgment and output personalities are reconstructed in accordance with the switching of value criteria. In this embodiment, the personality OS unit 210 may be configured to have a hierarchical structure including at least a value criteria layer 221 (top level), a responsibility and consistency layer 222, a judgment layer 223, and a personality layer 224 (output personality). This allows for the stable determination of response policies and output personalities as a "personality OS-like" control system centered on higher-level control based on value criteria, rather than simply maintaining a fixed persona description.

[0110] The Value Criteria Layer 221 is the highest-level control layer that defines the fundamental values ​​of the entire personality. The Value Criteria Layer 221 defines, for example, base values ​​such as integrity, integrity, and responsibility, and these definitions are used as control criteria that define the behavior of the lower layers (Responsibility / Integrity Layer 222, Judgment Layer 223, and Personality Layer 224). The Value Criteria Layer 221 may be managed by a Value Criteria Management Unit 212 (which may be omitted from the illustration), and the Value Criteria Management Unit 212 may maintain at least one Value Criteria Profile (for example, 221a and 221b in Figure 5).

[0111] The Responsibility and Consistency Layer 222 is a layer equipped with consistency checking processing to evaluate the consistency between the Value Criteria Layer 221 and the lower layers (Judgment Layer 223 and Personality Layer 224) and to reconstruct them without contradiction. The Responsibility and Consistency Layer 222 may, for example, refer to (a) the base values ​​or value criterion profile defined by the Value Criteria Layer 221, (b) the judgment logic or priority rules adopted by the Judgment Layer 223, and (c) the output personality parameters (politeness, degree of avoidance of assertion, tone of voice, etc.) set by the Personality Layer 224, and evaluate whether there are any inconsistencies or contradictions among these three. If a lack of consistency is detected, the Responsibility and Consistency Layer 222 may be configured to converge to a personality state consistent with the value criteria by modifying the parameters or applicable rules of the Judgment Layer 223 or the Personality Layer 224 (see the dotted arrow in Figure 5).

[0112] The judgment layer 223 is a layer that constitutes the judgment logic related to response policy or response selection, taking into consideration the output (or constraints) of the value criteria layer 221 and the responsibility / consistency layer 222. The judgment layer 223 functions, for example, as a judgment mechanism that concretizes the response policy determination (process S49) in UERA, and may decide which response policy to prioritize by taking into consideration the intent interpretation result, the sentiment recognition result, and the responsibility mapping result (9-direction responsibility vector). The response policy here may include at least output policies (normal response / unclear response / additional question / warning, etc.), safety-related policies (safety priority, topics to avoid, forbidden word strength, etc.), expression policies (politeness, strength of assertion avoidance, ambiguity, tone, presence or absence of empathetic preamble, etc.), permission / prohibition designation, and template designation (template identifier).

[0113] The personality layer 224 is a layer that defines the output personality (personality state at the time of output) based on the response policy determined by the judgment layer 223. The personality layer 224 may output parameters including, for example, the difference layer applied to the persona unit 216 by the PPI 214 (see Figure 7), the specification of permitted / forbidden actions and compatibility layers, and style differences and constraint strengths (politeness, assertion avoidance strength, etc.) that are passed to the syntactic control type response generation unit 204. This makes it possible to stably reflect an output personality that is consistent with the value criteria at the generation stage while maintaining the same outer frame (processes S1 to S7 in Figure 2).

[0114] The value criteria layer 221 maintains multiple value criteria profiles (e.g., 221a, 221b), and by switching between value criteria profiles, it is possible to replace the values ​​with those reflecting the values ​​of a different personality or individual. Switching between value criteria profiles may be done based on operator specification, instructions included in user input, or internal triggers (e.g., change in phase ID, change in relational information, increase in responsibility level). Value criteria profiles can be implemented as data containing a set of base values ​​(e.g., integrity, consistency, responsibility) and their priority or interpretation rules, and the data format may be any structured format.

[0115] In response to a change in the value criteria profile, the judgment logic of the judgment layer 223 (response tendencies, priority rules, weighting, etc.) and / or the output personality parameters of the personality layer 224 (tone of voice, speaking tendencies, preferred emotional tone, politeness, assertion avoidance strength, etc.) may be automatically readjusted, and the personality structure may be reconstructed. In other words, "change in value criteria → reconstruction of judgment and personality" operates as a single unit. Furthermore, in this embodiment, the consistency check process of the responsibility and consistency layer 222 ensures that the value criteria layer 221, judgment layer 223, and personality layer 224 are reconstructed without contradiction. For example, if a contradiction is detected where an action category that is not acceptable in light of the value criteria is permitted on the personality layer side, or where strong assertive expressions are preferred contrary to the assertion avoidance required by the value criteria, the responsibility and consistency layer 222 may modify the parameters or application rules of the judgment layer 223 or personality layer 224 to converge to a consistent policy and output personality.

[0116] As described above, by providing the value criteria layer 221 as a replaceable top-level control layer, and by configuring the lower layers (judgment layer 223 and personality layer 224) to be reconfigured in accordance with the switching of value criteria, and by ensuring that the reconfiguration is consistent through the consistency check of the responsibility and consistency layer 222, this system can function as a personality OS that reconfigures the personality structure and generates responses without relying on fixed personality settings.

[0117] The personality OS unit 210 (see Figure 3) interprets the situation of the dialogue using a responsibility vector (see Figure 6), which represents the situation as a "mathematical vector with a direction of responsibility," and can perform responsibility mapping processing or response selection processing based on this responsibility vector. Here, "direction of responsibility" means quantifying, as multiple directional components, the fact that even with the same utterance, the priority axis (slope of responsibility) in the response may differ, such as situations where empathy should be prioritized, situations where logical consistency should be prioritized, or situations where safety should be prioritized.

[0118] In one example provided herein, the responsibility vector R includes components in nine directions. For example, it may be expressed as follows (this is illustrative and not limiting): Equation (1) R = (r1, r2, r3, r4, r5, r6, r7, r8, r9) Each component r1 to r9 is a score representing the strength of responsibility in each direction within the dialogue situation. The names of the nine directions can be, for example, as follows: r1: Responsibility for emotional consideration r2: Honesty and Rational Responsibility r3: Support and Responsibility r4: Freedom, Intuition, and Responsibility r5: Order and Logical Responsibility r6: Responsibility for self-reflection and integration of contradictions r7: Emotional margin responsibility r8: Transformation and Possibility Responsibility r9: Protection, resolve, responsibility

[0119] The names listed above are examples of labels used to distinguish responsibility directions based on the classification principle of the nine-direction responsibility vector, and are not limited to the names themselves. For example, it is permissible to refer to equivalent concepts with different terms, or to slightly adjust the scope of equivalent concepts according to operational requirements (e.g., swapping specific elements included in the same direction). However, it is desirable that the responsibility vector represents "responsibility in multiple directions" and includes components in all nine directions (even when integrating or splitting directions, it should be mapped in a way that allows it to be treated as nine-direction components).

[0120] The calculation of each component r_i is not limited, but from the standpoint of feasibility requirements, it may be implemented as follows, for example. First, m features x_1 to x_m are extracted from the input (user utterance), dialogue context, relationship information (such as the phase ID held by the relationship log management unit 202), and analysis results (such as responsibility depth and sentiment score), and a feature vector x is generated. Equation (2) x = (x_1, x_2, …, x_m) Here, each feature x_j may be extracted using a rule-based method, or it may be obtained as the output (probability / score) of a classifier or regressor. Next, for each direction i (i=1~9), the unnormalized score s_i (raw score before normalization) is calculated using the weights w_{i,j} and the bias b_i. For example, it may be calculated by a linear combination as shown in the following equation (this is an example and not limited to this). Equation (3) s_i = Σ_{j=1}^{m} (w_{i,j}·x_j) + b_i (i=1~9) Here, Σ_{j=1}^{m} represents the sum from j=1 to m, w_{i,j} is the weight of the j-th feature in the i-th direction, and b_i is the bias. w_{i,j} and b_i may be determined through learning or manually set according to operational rules.

[0121] Each component r_i is obtained by normalizing s_i to a predetermined range. For example, to normalize to 0 to 1 and satisfy Σ_{i=1}^{9} r_i = 1, softmax normalization may be used (this is an example and not a limitation). Equation (4) r_i = exp(s_i) / Σ_{k=1}^{9} exp(s_k) (i=1~9) In this case, r_i is in the range of 0 to 1, and Σ_{i=1}^{9} r_i = 1.

[0122] Other normalization methods that may be used include min-max normalization or sigmoid transformation. For example, an example of min-max normalization is given by the following equation (this is illustrative and not limiting): Equation (5) r_i = (s_i - s_min) / (s_max - s_min + ε) Here, s_min = min(s_1~s_9), s_max = max(s_1~s_9), and ε is an infinitesimal value to avoid division by zero. Alternatively, we can map from 0 to 1 using a sigmoid function as shown in the following equation. Equation (6) r_i = 1 / (1 + exp(-s_i)) Furthermore, the responsibility vector may be time-smoothed not only to ensure stability in continuous dialogue, but also as a result of a single turn. For example, the responsibility component r_i(t) at time t may be updated using the previous value r_i(t-1) and the newly calculated value r_i,new(t) with the following formula (this is an example and not an limitation). Equation (7) r_i(t) = (1 - λ)·r_i(t-1) + λ·r_i,new(t) Here, λ is a smoothing coefficient between 0 and 1, and r_i,new(t) are values ​​newly calculated using equations (4) to (6), etc.

[0123] The calculated responsibility vector R can be used as input for determining the response strategy. For example, the direction corresponding to the largest component may be adopted as the preferred responsibility direction. Equation (8) i* = argmax_{1≦i≦9} r_i Based on this i*, the response strategy (e.g., strength of avoiding definitive statements, level of politeness, level of detail in explanations, priority of safety templates, etc.) may be adjusted. Furthermore, if multiple directions are to be considered simultaneously, multiple candidate strategies may be weighted and combined using r_i as the weight.

[0124] The nine-directional responsibility vectors described above can be used for responsibility mapping processing, responsibility frame determination, responsibility assessment in the Self-Correcting Engine (SRPE), and fitness calculation in multi-personality resonance (see Figures 4, 8, and 9). This allows for a quantitative representation of responsibility bias (which responsibilities should be prioritized) depending on the context of the dialogue, and can be reflected in the decision-making process for response strategies or personality settings.

[0125] Figure 7 is a block diagram showing the separation structure of the core unit 215 and the persona unit 216 in the personality OS unit 210 (see Figure 3), and an example of a personality plug-in interface (PPI) 214 connecting the two. In this embodiment, the personality is defined as being separated into a "core unit 215" which is responsible for control over responsibility, values, and introspection, and a "persona unit 216" which is responsible for tone, speech patterns, and character attributes. The persona unit 216 can be applied via a PPI 214 that allows connection to the core unit 215. This allows for independent updating and switching of higher-level control (core) based on value standards and responsibility vectors, etc., and character traits on output (persona), enabling both personality consistency and operational flexibility in long-term dialogue.

[0126] The core unit 215 is the central hub for determining response policies based on responsibility, values, and introspection, and includes at least value criteria information (see Figure 5), responsibility mapping (see Figure 6), and introspection (see process S46 in Figure 4). The core unit 215 generates a policy record that includes the response policy (output policy, safety-related policy, expression policy, permission / prohibition, template specification, etc.) determined in process S49 in Figure 4, and may pass this policy to the prompt reconstruction unit 201 and the syntactically controlled response generation unit 204, etc. In contrast, the persona unit 216 is mainly responsible for the tone, speech style, and character attributes of the output sentence (e.g., sentence ending tendencies, vocabulary preferences, expression style of honorific hierarchy, form of self-referentiality, explanatory style, etc.), and may concretize the output style to the extent that it is consistent with the policy determined by the core unit 215.

[0127] PPI214 is an interface for passing persona differences in an applicable format between the core unit 215 and the persona unit 216. PPI214 may consist of at least (a) a persona core, (b) guiding motto information, (c) a difference layer, (d) permitted / forbidden actions, and (e) a compatibility layer. Here, the persona core is information that defines the basic identity of the persona, and may include, for example, a display name, self-referential terms, rules for addressing users, standards for basic writing style (polite / informal, etc.), and standards for explanatory policies. The guiding motto information is a collection of short mottos or principle sentences that serve as guidelines for the persona, and may include, for example, principle sentences such as "Be honest, prioritize safety." The difference layer represents the changes (differences) to the persona core, and may be expressed by tags + weights, rule sets, or template identifiers, as described below. The permitted / prohibited actions specify the categories of actions to be allowed and prohibited in the output, and can be applied in a manner consistent with the value standards and safety policies of the core section 215. The compatibility layer is a layer of transformations and mappings that makes persona differences applicable between different large-scale language models (LLMs), and, as described below, it acts as a bridge between model-independent and model-dependent representations.

[0128] From a feasibility standpoint, the representation of the difference layer may be implemented in at least one of the following ways (or a combination thereof): As a first example, the difference layer may be represented as a set of "tags + weights," specifying the direction of the output style quantitatively or semi-quantitatively, for example, "politeness +2," "assertiveness -1," "empathy preface +1," "metaphorical expression +0," etc. As a second example, the difference layer may be represented as a "rule set," implemented as a set of constraints, for example, "polite endings," "add buffering expressions when negative," "avoid dangerous vocabulary and replace with alternative words," "prioritize bullet points," etc. As a third example, the difference layer may be represented as a "template identifier (template ID)," specifying, for example, a combination of a response template, an explanation template, or a safety template. These difference layers may be applied as slot selection and priority control (insertion / removal from internal prompts) in the prompt reconstruction unit 201, or as parameters such as lexical bias, regular expression constraints, and finite state transducer (FST) constraints in the syntactically controlled response generation unit 204 and the generation constraint application 302.

[0129] Permitted / prohibited actions are specifications that control the range of output actions that a persona may take. For example, permitted actions may include "general advice," "presenting additional questions," "presenting options," "warnings," and "expressing consideration for the user's feelings," while prohibited actions may include "unfounded assertions," "encouraging risky behavior," "presenting inappropriate instructions," and "leading the user to violate their privacy." Permitted / prohibited actions are set to be consistent with the value criteria information (see Figure 5) and safety policies (response policy flags, etc.) in the core section 215, and in the event of a conflict, the value criteria or safety policy may take precedence. This allows for the suppression of outputs that are unacceptable to the system while maintaining the persona's style.

[0130] The compatibility layer is a layer for defining differences related to the persona unit 216 in a format applicable to the core unit 215, even when selecting or switching between multiple types of large-scale language models that generate response sentences (see LLM inference 301 in Figure 1). In this embodiment, the compatibility layer holds the difference layer and permission / prohibition actions as model-independent intermediate representations (e.g., tags + weights, abstract rules, template identifiers, etc.), and may convert and apply them to model-dependent representations such as slot indications in the prompt reconstruction unit 201, style, vocabulary, and syntactic rules in the syntactic-controlled response generation unit 204, and lexical biases and regular expressions / FST constraints in the generation constraint application 302, depending on the selected LLM. This ensures that persona differences (tone, speech, and attributes) can be consistently applied even when the LLM is switched, and that personality continuity and compatibility are ensured.

[0131] The persona section 216 defined by PPI 214 is applied in a manner consistent with the response policy determined by the core section 215 (see process S49 in Figure 4). That is, if the core section 215 sets a high "high responsibility flag" or "assertion avoidance strength," the compatibility layer may convert and apply the data via the difference layer to suppress assertive expressions, prioritize additional questions, or apply safety templates. Conversely, if there is a strong intention for creation or exploration and a degree of freedom is allowed, the difference layer may adjust the data to increase vocabulary preferences for ideation support and allow for figurative expressions. These applications may be configured to act on the probability distribution at the generation stage (syntactic-controlled response generation section 204 and generation constraint application 302) rather than simply converting the text at a later stage.

[0132] As described above, by separating the core section 215 and the persona section 216, and by making it possible to apply persona differences via PPI 214 (persona core / guideline slogan information / difference layer / permitted / prohibited behavior / compatibility layer), it is possible to maintain higher-level control (core) based on value standards, responsibility, and introspection while ensuring compatibility when switching, updating, and switching LLMs for output personalities (personas).

[0133] Figure 8 is a flowchart showing an example of a Self-Refinement Proposal Engine (SRPE) that the personality OS unit 210 (see Figure 3) may have. The self-refinement engine is a processing system that allows the conversational AI to evaluate its own response history or internal state, and if it detects inconsistencies or contradictions with related information and personality settings, it proposes corrections to internal prompts or readjusts the response policy, and reflects these in subsequent processes (such as process S4 and / or process S5 in Figure 2). The self-refinement engine may be executed in conjunction with the response output (process S6 in Figure 2), or it may be executed at predetermined timings (e.g., every certain number of turns, when the phase ID is updated, when the differential indicator changes abruptly, etc.).

[0134] The self-correction engine may consist of at least three modules: (a) a contradiction detection module 241 that detects inconsistencies or contradictions, (b) an improvement proposal generation module 242 that generates a correction policy or proposed correction, and (c) a reflection / learning module 243 that reflects the proposed correction in internal prompts or response policies for learning (see Figure 8). This allows for the integration of not only detection but also the generation and reflection of proposed corrections, enabling the maintenance of personality consistency without relying on manual intervention by the operator.

[0135] The process in Figure 8 can be executed in the order of steps S81 to S89, as shown in the figure. In step S81, the inconsistency detection module 241 detects inconsistencies or contradictions. The detection targets may include at least the most recent response, the most recent internal prompt (or its summary), relationship information held in the relationship log management unit 202, personality settings output by the personality OS unit 210 (which may include the specification of PPI 214), and value criteria information (see Figure 5). The inconsistency detection module 241 may calculate a detection index (score, flag, or composite index) indicating the inconsistency between these.

[0136] In step S82, it is determined whether or not an inconsistency exists based on the detection result of step S81. If no inconsistency exists (step S82 is negative), the self-correction engine processing may be terminated. If an inconsistency exists (step S82 is positive), the following steps S83 onwards are executed to propose corrections to the internal prompt or readjust the response policy.

[0137] Process S83 involves conducting a safety check. This safety check may include detecting hazardous language, verifying compliance with ethical guidelines (or operational policies), and determining the appropriateness of avoiding definitive statements in high-responsibility areas. The results of the safety check may be used as constraints in subsequent revision generation (e.g., strengthening prohibited behaviors, prioritizing safety templates, recommending expert consultation, etc.).

[0138] In step S84, a responsibility frame determination is performed. Responsibility frame determination may be a process that takes into account responsibility depth, nine-directional responsibility vectors (see Figure 6), and relational information (phase ID, difference indicator, emotional events, etc.) to determine the responsibility direction or safety priority that should be prioritized in the current response (or the next response). The results of the responsibility frame determination may be reflected in subsequent personality style reflection (step S85) and response selection (step S87).

[0139] In step S85, personality style reflection is performed. Personality style reflection may be a process that evaluates whether there is a style deviation in the most recent response in light of the persona differences (difference layer, permitted / forbidden behavior, compatibility layer, etc.) specified in PPI214 (see Figure 7) and the response policy determined by the personality OS unit 210 (politeness, strength of avoiding assertions, template specification, etc.), and adjusts the strength of style constraints or application rules for subsequent responses as necessary. For example, the level of honorifics, sentence ending tendencies, naming rules, forbidden vocabulary, and strength of avoiding assertions may be adjusted.

[0140] In step S86, emotional tone adjustment is performed. Emotional tone adjustment may be a process that adjusts the expression policies such as reassurance preambles, empathy expressions, ambiguity, and emphasis, taking into consideration the emotional score value calculated by the analysis unit 203, the presence or absence of emotional events (see step S3 in Figure 2), and related information. Emotional tone adjustment is performed in a manner consistent with the reflection of personality style, and may include rules such as increasing considerate expressions when the emotional score is high, and avoiding excessive assertions of reassurance when the responsibility level is high.

[0141] In step S87, a response selection is performed. Response selection may involve determining the preferred response policy (e.g., normal response / unclear response / additional question / warning / provision of action / recommendation of expert consultation, etc.) or template designation for the next response, taking into account the constraints and adjustment results obtained in steps S83 to S86. If necessary, priority rules may be applied, such as prioritizing an unclear response when the confidence score is low, or prioritizing a safer policy (provision of action or recommendation of consultation, etc.) when the level of responsibility is high.

[0142] In step S88, the improvement proposal generation module 242 generates a correction policy or a proposed correction. The proposed correction may include at least a suggested correction of the internal prompt, a proposed readjustment of the response policy, and a proposed adjustment of the generation constraints. For example, the proposed correction may include adding prohibited vocabulary, increasing the assertion avoidance strength, switching template identifiers, adjusting the difference layer weights of the PPI, or changing the safety priority (constraint priority). The improvement proposal generation module 242 may switch the type of proposed correction to be generated depending on the detection indicator (what type of inconsistency) of the inconsistency detection module 241.

[0143] In process S89, the reflection / learning module 243 reflects the proposed modifications generated in process S88 into the internal prompt or response policy and provides it for learning. "Reflection" here may include, for example, (a) modifications to slots (or instructions) to be inserted / removed in the next process S4 (prompt reconstruction), (b) changes to the priority or strength of the rule set (regular expression, FST, lexical bias, etc.) to be applied in the next process S5 (syntactic-controlled response generation), (c) updates to the difference layer or permitted / forbidden actions in PPI 214, (d) updates to thresholds (e.g., thresholds for safer branches) or weight coefficients, and (e) addition of modification events to the relationship log management unit 202. The reflection / learning module 243 may work in conjunction with the learning logic 205 to continuously calibrate the effects of the modifications, taking into account explicit or implicit evaluations from the user.

[0144] Specific examples of "contradictions and inconsistencies" that the contradiction detection module 241 may detect include the following (these are examples and not limiting): (a) Logical contradiction: Making mutually incompatible claims within the same response or between consecutive turns, or inconsistencies between premises and conclusions. (b) Style deviation: Outputting expressions that violate the honorific hierarchy, ending tendencies, naming rules, or prohibited vocabulary specified by PPI 214. (c) Value standard inconsistency: Including excessive assertions, unfounded recommendations, or a lack of safety considerations that contradict the integrity, consistency, and responsibility required by the value standard layer 221 (see Figure 5). (d) Discrepancy with relational information: Displaying names, attitudes, or policies that contradict relational information (roles, phase IDs, known agreements, etc.) held by the relational log management unit 202. (e) Responsibility frame inconsistency: Prioritizing highly flexible suggestions or assertive expressions even in situations where a safer approach should be taken in light of the depth of responsibility or the nine-way responsibility vector.

[0145] As described above, the Self-Correction Engine (SRPE) can perform processing in the following order: inconsistency detection (241) → safety check → responsibility frame determination → personality style reflection → emotional tone adjustment → response selection → improvement proposal generation (242) → reflection and learning (243). This allows it to go beyond single-point "error detection" and propose and reflect corrections in a manner consistent with value standards, responsibility, emotions, and relationships, thereby improving personality consistency, safety, and operational stability in long-term dialogues.

[0146] Figure 9 is a block diagram showing an example of a "Multi-Persona Resonance" unit 218 in the personality OS unit 210 (see Figure 3) that simultaneously holds and coordinates multiple persona units. In this embodiment, the system is not limited to a configuration that applies only a single persona unit 216, but can simultaneously hold multiple persona units and decide which persona to select as the responder or whether to coordinate multiple personas depending on the input / context. This allows for appropriate switching and integration of different discourse styles such as explanatory, empathetic, and reflective support, depending on the purpose and direction of responsibility of the dialogue (see Figure 6). Furthermore, it may be implemented as a conversational AI in which multiple persona units cooperate collectively, and such a configuration is illustrative and does not limit the invention of this disclosure.

[0147] As shown in Figure 9, the multi-personality resonance unit 218 may be configured to include at least (i) a personality context pool that holds the state of each of the multiple persona units, (ii) a resonance filter that evaluates the degree of fit of each persona unit to the input, (iii) a personality selector that determines the persona unit responsible for the response based on the degree of fit, and (iv) a consistency maintenance engine that controls the response of the selected persona unit so that it is consistent with the control of responsibility, values, and introspection in the core unit 215. Figure 9 shows an example in which input information (context / relationship information / responsibility, emotions, etc.) is input to the personality context pool, passes through the resonance filter, personality selector, and consistency maintenance engine, outputs a selected persona, and hands over the selected persona to the PPI 214 (see Figure 7).

[0148] The personality context pool may be configured to hold at least the responsibility vector, the most recent emotional state, and the content of the previous response for each persona. The responsibility vector may include the nine-directional responsibility vector (or a summary thereof) shown in Figure 6. The most recent emotional state may include the emotional score value from the analysis unit 203, the presence or absence of an emotional event, or the emotional state label derived therefrom. The content of the previous response may be held as the most recent output sentence itself within the scope of the "latest part" (full text retention window) in the abridgement and forgetting unit 206, and after the output sentence moves to the "past part," it may be held as a summary generated by the abridgement and forgetting unit 206, or as style features (e.g., politeness, avoidance of assertion, sentence ending tendency, etc.). Furthermore, the personality context pool may hold references to relationship information (user identifier, phase ID, relationship depth, difference index, etc.) held in the relationship log management unit 202. This makes it possible to refer to the "current state" of each persona, including across sessions.

[0149] The resonance filter evaluates the degree of fit for each persona to the input. The degree of fit may be calculated, for example, based on the input intent (see process S43 in Figure 4), the nine-direction responsibility vector (see Figure 6), the emotion score value calculated by the analysis unit 203 based on the input, and the relationship information held in the relationship log management unit 202. The calculation method is arbitrary, but for example, for each persona p, the degree of fit may be obtained by weighted synthesis of (a) the closeness between the input responsibility vector and the persona's preferred direction vector, (b) the fit between the emotion score value and the persona's expression policy (empathy-oriented / explanatory, etc.), and (c) the fit with relationship information (phase ID, role, intimacy, etc.). The resonance filter may output the degree of fit as a numerical score or as a group of top candidates (candidate set).

[0150] The personality selector may select a single persona, or it may be configured to select multiple personas and perform a collaborative response. The collaborative response may include, for example, (a) a mode in which response proposals or perspectives generated by multiple personas are integrated in the same turn and output as a single response sentence with consistency ensured by the consistency maintenance engine, and / or (b) a mode in which the main persona is switched over multiple turns while the continuity of titles, honorific levels, sentence endings, etc. is maintained by the PPI214 and the consistency maintenance engine. In the case of multiple selections, for example, a main persona and a sub-persona may be designated, with the main persona responsible for the basic tone and structure, and the sub-persona providing additional perspectives (logical reinforcement, empathy reinforcement, introspection reinforcement, etc.). As an example of multiple personas, one could simultaneously maintain Persona A, who emphasizes empathy, and Persona B, who emphasizes logical structuring, and then select or coordinate them according to the input intent and direction of responsibility (this is just one example, and the persona names, number, and roles are not limited).

[0151] The consistency maintenance engine controls the responses of the selected persona unit to be consistent with the controls related to responsibility, values, and introspection in the core unit 215. In other words, the consistency maintenance engine may impose constraints on the persona differences (difference layer in Figure 7, permitted / forbidden actions, etc.) to prevent deviations from the value criteria information (see Figure 5) and response policies (politeness, assertion avoidance strength, permitted / forbidden actions, template specification, etc.). Furthermore, even when multiple persona units respond alternately or in cooperation, the response policies (sentence endings, honorific hierarchy, addressing rules, assertion avoidance strength, safety priority, etc.) may be uniformly adjusted so that the whole behaves as a consistent personality. This maintains consistency as a conversational AI observed by the user, even when multiple personalities are used.

[0152] As shown in Figure 9, the selected persona (or group of personas) may be configured to be applied via PPI214 (see Figure 7). That is, the output of the personality selector or consistency maintenance engine may be passed to PPI214 as a persona identifier, specification of the difference layer, specification of permitted / forbidden actions, and specification of the application of the compatibility layer, etc. Based on these specifications, PPI214 may apply the differences to the prompt reconstruction unit 201 and the syntactic control type response generation unit 204 (and generation constraint application 302) in an implementable manner.

[0153] Furthermore, the multi-personality resonance unit 218 may also be configured to include a hub (resonance hub) that oversees the processing of the personality context pool, resonance filter, personality selector, and consistency maintenance engine. The resonance hub may be responsible for, for example, controlling the processing order, managing parameters for fitness calculation, managing the order of statements (statement queue) when multiple personalities cooperate, and suppressing bias in relational information (excessive fixation to a specific persona). In Figure 9, the resonance hub is omitted for simplification of the diagram, but in implementation, providing a processing system that oversees these will allow for flexible adaptation to future functional divisions and corrections (e.g., increasing or decreasing the number of personas, adding evaluation rules, etc.).

[0154] As described above, the multi-personality resonance unit 218 can be configured to calculate the degree of fit based on responsibility vectors, emotion score values, and relationship information, select one or more personas, and adjust and output them in accordance with the control of the core unit 215. This allows for the concrete implementation of the multi-personality resonance configuration (at least a personality context pool, resonance filter, personality selector, and consistency maintenance engine), enabling both personality consistency and situational adaptability in long-term dialogues.

[0155] Figure 10 is a flowchart showing an example of an Emotion Weighted Memory Filtering Selector (EW-MFS) that the personality OS unit 210 (see Figure 3) may have. The processing in Figure 10 may be performed in conjunction with step S7 (Collapsed Forgetting) of the main flow shown in Figure 2, or it may be performed as an internal process of step S7. In this embodiment, a memory priority (E score) is calculated for events (experiences / incidents) that occur during a conversation, and the event is classified into at least one of the following: actual memory, temporary memory, compressed memory, and discarded, and then integrated at a predetermined timing to organize and store as an episode.

[0156] In this specification, "event" refers to an information unit representing an event that occurred during a conversation, and may include, for example, user input, AI response, explicit or implicit evaluation by the user, sentiment event, phase ID update, policy switch, or correction suggestion by the self-correction engine (see Figure 8). Events may be treated in association with metadata such as timestamp, user identifier, phase ID, responsibility depth, sentiment score value, and difference indicator. While the event log 401 held by the relationship log management unit 202 may be maintained as an append-only log for auditing purposes, "discarding" in EW-MFS may mean excluding it from storage in long-term memory (e.g., summary store 402 and feature store 403, etc.) for future response policy decisions, and does not prevent recording or retention (for auditing purposes) in the append-only event log 401, nor does it include deletion or erasure of the event log 401.

[0157] In step S101 of Figure 10, EW-MFS acquires events. The source of events is arbitrary, but for example, events extracted from the dialogue text (original text), relationship information recorded by the relationship log management unit 202, analysis results by the analysis unit 203 (depth of responsibility, emotion score, confidence level, etc.), and response policies or personality settings by the personality OS unit 210 (may include PPI application results) may be taken into consideration when generating or extracting events for memory selection. The acquired events may be normalized into an internal representation that can be referenced in subsequent steps.

[0158] In step S102, EW-MFS calculates the memory priority (E score). In this embodiment, the E score is calculated using at least empathy, importance, and responsibility depth as elements, and at least one of relationship depth, responsibility priority, and contextual importance can be optionally considered. Here, empathy is an index representing the impact of the event on the user's emotional state or the need for empathetic consideration, and may be derived based on the emotional score value of the analysis unit 203, the presence or absence of an emotional event, and the emotional weight of the event (e.g., strong anxiety, joy). Importance is an index representing the contribution of the event to task achievement, relationship formation, or future response policy determination, and may be derived based, for example, on the user's explicit important instructions, repeated mentions, milestones in phase ID (achievement, agreement, decision, etc.), or impact on subsequent branches (unclear response, recommendation to consult, etc.). Responsibility depth may utilize the responsibility depth calculated by the existing analysis unit 203 (see Figure 2). Relationship depth may be derived based on relationship information (proxy indicators of intimacy, difference indicators, etc.) held in the relationship log management unit 202. Responsibility priority may be derived based on the priority direction in the 9-direction responsibility vector (see Figure 6) or the result of responsibility frame determination (see Figure 8). Contextual importance is the importance of the event in question in the current context (phase ID, topic, situation), and may be derived based on, for example, proximity to the session objective, the need to avoid serious misunderstandings, or contribution to safety. Furthermore, EW-MFS may optionally consider a topic diversity index based on the degree of topic dispersion in order to suppress memory bias of events or episodes. The topic diversity index can be calculated based on, for example, the number of types of phase IDs, the distribution of topic tags, or the number and dispersion of clusters of embedded representations. If the topic diversity index is lower than a predetermined threshold, the retention priority (E score or classification threshold) of events other than the topic in question may be relatively corrected upward, or a correction may be made to suppress the retention ratio of events of the topic in question, in order to suppress memory concentration on the same topic. Furthermore, EW-MFS may optionally calculate a Time Zone Balance index that represents the bias in the timing of inputs or events, based on timestamps and time zone information associated with the user. The Time Zone Balance index is, for example, a quantity that represents the bias in the input distribution (morning / noon / evening / night, etc.) or daily boundaries in the user's local time, and the execution time of the integration (at a predetermined timing) described later, or the weighting of the E score (for example, correction of the retention priority of high-emotion events that occurred during the nighttime) may be adjusted according to this index.

[0159] The method for calculating the E score is arbitrary, but from a practical standpoint, it may be done as follows: After normalizing each element to the range of 0 to 1, it may be calculated using a weighted linear combination as follows (if no arbitrary elements are used, the corresponding weight may be set to 0). E =(w_e×E_empathy)+(w_i×E_importance)+(w_r×E_resp)+(w_rel×E_relation)+(w_pri×E_priority)+(w_c×E_context) Here, E_empathy may represent the degree of empathy, E_importance the degree of importance, E_resp an index derived from the depth of responsibility, E_relation the depth of the relationship, E_priority the priority of responsibility, and E_context the contextual importance. The coefficients may be normalized to satisfy w_e+w_i+w_r+w_rel+w_pri+w_c=1. Furthermore, to clarify that it is emotion-weighted, the configuration may dynamically increase or decrease w_e according to the degree of empathy (or emotion event) (e.g., increase w_e when the emotion score value exceeds a threshold). Note that the E score may also optionally consider at least one of the above elements, namely Topic Diversity and / or Time Zone Balance. In this case, weight coefficients corresponding to the said index may be added and linearly combined, and the coefficients as a whole may be normalized.

[0160] In step S103, the event is classified based on the calculated E-score. The classification categories include at least main storage, temporary storage, compressed storage, and discard. The classification rules are arbitrary, but for example, a threshold system may be adopted in which the higher the E-score, the higher the retention priority. Specifically, if the E-score is above the first threshold, it may be stored in main storage; if it is above the second threshold but below the first threshold, it may be stored in temporary storage; if it is above the third threshold but below the second threshold, it may be stored in compressed storage; and if it is below the third threshold, it may be discarded. The thresholds and weights may be calibrated based on operational data (explicit evaluation, implicit evaluation, etc.).

[0161] Main memory refers to high-priority events that should be directly referenced in future response decisions, and may include, for example, important agreements related to safety, assumptions for ongoing relationships, or important decisions in high-responsibility areas. Provisional memory refers to candidate events that do not need to be definitively retained at the present time, but whose importance may increase through repetition or subsequent developments, and may include, for example, minor preferences, initial trial remarks, or transient emotional reactions. Compressed memory refers to events that should be retained but not in detail, and should be retained in a compressed form as a summary or feature, and may include, for example, representative summaries when similar events occur frequently, or abstractions of lengthy events. Discarded memory refers to events that will not be retained as long-term memory, at least not for use in future response decisions, and may include, for example, noise, accidental greetings, or fragments that do not contribute to maintaining context.

[0162] In step S104, the process branches according to the classification result of step S103. If the classification result is for main storage, the process proceeds to step S106; if the classification result is for temporary storage or compressed storage, the process proceeds to step S105; and if the classification result is for discard, the process proceeds to step S107.

[0163] In step S105, events classified as temporary memory and / or compressed memory are integrated and organized into episodes. The timing of the integration (predetermined timing) is arbitrary, but from the standpoint of feasibility, it may be performed, for example, at the end of a session, when the phase ID changes, after a certain period of time has elapsed, or when the temporary memory buffer reaches a predetermined number of items. The unit of integration may be, for example, a group of events that occurred within a certain period of time under the same phase ID, or a group of events belonging to the same topic. The episode generated as a result of the integration may be a summary expression that includes the main points of the events, causal relationships, and features that should be referenced in the future (e.g., bias in direction of responsibility, emotional shifts, agreements, etc.). Another example of the predetermined timing may include a nightly review performed at a predetermined time such as the daily boundary in the user's time zone or during the late-night period. During the daily review, the E-score may be recalculated, the classification re-evaluated (e.g., promoted from temporary memory to permanent memory, or demoted to discarded), and episode consolidation may be performed on the event groups held in the temporary memory buffer and / or classified into compressed memory, while referring to the difference indicator for the period, the presence or absence of sentiment events, and the aforementioned topic diversity indicator or time-of-day balance indicator, etc.

[0164] In step S106, an event or episode is saved. The storage location may be the summary store 402 and feature store 403 of the data layer shown in Figure 1. For example, the summary store 402 may store an episode summary (or a summary of the main memory event), and the feature store 403 may store features used for future response strategy decisions (e.g., statistics on responsibility direction, sentiment transition, number of re-questions, surrogate indicators of relationship changes, etc.). Furthermore, the classification results (main memory / temporary memory / compressed memory / discard) and the E score itself may be added to the relationship log management unit 202 as structured metadata. Furthermore, as a result of the Nightly Review or the integration (process S105), a log record (e.g., a daily summary record) summarizing the dialogue or episode during the period (e.g., one day) may be generated and stored in the summary store 402. The log record may include, for example, key points of the day's main phase ID, important agreements, a summary of the emotional shifts, and excerpts of notes that will help determine the next response strategy. The log record may be generated in a form that reduces or anonymizes non-identifiable proper names or contact information in accordance with the personal information minimization policy.

[0165] In step S107, events classified as discard are discarded. Discarding means that they are not saved to at least the summary store 402 and the feature store 403. This prevents the long-term retention of information that is unnecessary for future response policy decisions. As mentioned above, this does not prevent the configuration of maintaining a minimum amount of additional information for auditing purposes (event log 401, etc.) separately.

[0166] To clarify the relationship (division of roles) between EW-MFS and the abridged forgetting unit 206 (see Figure 1), the abridged forgetting unit 206 is primarily responsible for "controlling the amount of memory of the dialogue text (summarization)," such as summarizing and retaining past portions of the dialogue text (original text) and retaining the most recent portions in their entirety. In contrast, EW-MFS is responsible for prioritizing (calculating E-scores) and classifying (main memory / temporary memory / compressed memory / discarding) events / experiences that occur during the dialogue, as well as episodicizing (integrating) temporary and compressed memories. In other words, abridged forgetting is primarily about controlling the retention format of the text (full text → summary), while EW-MFS is primarily about forming "referable semantic memories" through event-level importance judgment and episodicization. The two are complementary. For example, events that EW-MFS identifies as main memory may be retained as key points in summary generation during reduced forgetting, compressed memory may have a coarser summary granularity, and discarded items may be excluded from summarization. In this configuration, the summarization policy for reduced forgetting may be controlled by the classification results of EW-MFS.

[0167] As described above, the EW-MFS shown in Figure 10 can perform the calculation, classification, integration (episodicization), and storage (summary store 402 / feature store 403) of memory priority (E score) through processes S101 to S107. This allows the "meaning of the experience," which cannot be fully retained by simple summarization of the dialogue text alone, to be selected according to responsibility, emotion, and importance, and stored in a way that contributes to future response strategy decisions.

[0168] This section describes examples of how the personality OS unit 210, value-based profile switching and personality reconstruction (Figure 5), multi-personality resonance unit 218 (Figure 9), emotion-weighted memory selection unit (EW-MFS) (Figure 10), self-correction engine (SRPE) (Figure 8), etc., function as embodiments and contribute to personality consistency, safety, and operational stability in continuous dialogue. Note that the following are illustrative examples, and the inventions of this disclosure are not limited to these.

[0169] An example of how judgment and personality are restructured in response to a switch in value criteria profiles is explained (see Figure 5). In this embodiment, the value criteria management unit 212 maintains at least two value criteria profiles 221a and 221b. For example, value criteria profile 221a may include a policy that strongly prioritizes honesty, consistency, and responsibility, and emphasizes avoiding definitive statements and asking safe additional questions. On the other hand, value criteria profile 221b may include a policy that prioritizes empathy and increases reassuring preambles and empathetic expressions (these are examples and not limitations).

[0170] For example, consider a case where the same user requests "fact-checking and decision-making support" in one phase ID and "emotional support" in another phase ID. After input reception, analysis, and relationship log updating are performed in steps S1 to S3 (see Figure 2), the personality OS unit 210 determines the response policy in steps S41 to S49 shown in Figure 4. At this time, in step S42, the value standard management unit 212 may select (or switch) either value standard profile 221a / 221b in response to related information (changes in phase ID, differential indicators, past evaluation trends, etc.) or explicit instructions from the user.

[0171] When the value criteria profile is switched, as shown in Figure 5, the higher-level control of the value criteria layer 221 changes, and the judgment layer 223 and personality layer 224 are automatically readjusted (reconfigured) via the responsibility and consistency layer 222. For example, if value criteria profile 221a is selected, the judgment layer 223 may be reconfigured to prioritize avoiding assertions and asking follow-up questions, and the personality layer 224 may be reconfigured to increase politeness and safety template preference. On the other hand, if value criteria profile 221b is selected, the judgment layer 223 may be reconfigured to prioritize adding empathetic preambles and emotional tone, and the personality layer 224 may be reconfigured to increase warmth and considerate expression.

[0172] During this restructuring, the responsibility and consistency layer 222 performs consistency checks and may make modifications to ensure that no inconsistencies arise between the value criteria layer 221, the judgment layer 223, and the personality layer 224. For example, if a contradiction arises where the personality layer prefers strong assertive language in a situation where the value criteria require safety as the priority, the responsibility and consistency layer 222 may restore consistency by increasing the assertion avoidance strength, switching the template specification to the safety side, or modifying the permission / prohibition specification.

[0173] As a result, judgments and personality can be reconstructed in accordance with the switching of value criteria profiles, without relying on fixed personality settings, and response policies can be determined while suppressing contradictions. This allows for maintaining personality consistency that is appropriate to the context (phase) in continuous dialogue with the same user, while prioritizing and applying safer response policies when necessary.

[0174] This section describes an example of how the multi-personality resonance unit 218 (Figure 9) and the emotion-weighted memory selection unit (EW-MFS) (Figure 10) work together to improve the continuity of personalities in continuous dialogue. First, the multi-personality resonance unit 218 simultaneously holds multiple persona units and may select a single persona to respond depending on the input / context, or it may select multiple personas to respond collaboratively (see Figure 9). The personality context pool holds at least the responsibility vector, most recent emotional state, and previous response content for each persona. The resonance filter calculates the degree of fit based on the input intention, responsibility vector, emotion score value, relational information, etc., and the personality selector determines the persona(s) responsible for the response (single or multiple). The consistency maintenance engine maintains overall consistency among multiple personas to be consistent with the value standards, responsibility, and introspection of the core unit 215.

[0175] Next, EW-MFS (Figure 10) calculates an E-score (memory priority) for events (experiences) that occur during the conversation, classifies them into main memory / temporary memory / compressed memory / discarded, and integrates the events classified as temporary memory and / or compressed memory at predetermined timings to create episodes, which are then stored in the summary store 402 and the feature store 403 (steps S101-S107). The E-score is calculated using at least empathy, importance, and depth of responsibility as elements, and can also optionally consider relationship depth, responsibility priority, contextual importance, etc.

[0176] For example, if a user continues to consult on the same topic multiple times, events within each session (important agreements, strong emotional changes, policy shifts, etc.) will have a high E score and may be retained as main memory or compressed memory. On the other hand, casual conversational fragments or fragments that do not contribute to maintaining context will have a low E score and may be classified as discarded. Furthermore, at a predetermined timing such as the end of a session (process S105), the temporary memory / compressed memory is integrated and organized into an episode.

[0177] This episode and its features are referenced in the relationship log management unit 202 and the personality context pool in the next session and used for evaluating the fit of the resonance filter (Figure 9) and for determining the policy of the personality OS unit 210 (Figure 4). This allows important events (main memories / episodes) to be continuously referenced without having to retain the entire past text indefinitely, suppressing unnecessary fluctuations in the selected persona and expression policy. Therefore, even with a multi-personality configuration, overall personality continuity is improved, and consistency (continuity of tone, policy, and relationships) can be maintained across sessions.

[0178] This section explains an example of how the Self-Correction Engine (SRPE) (Figure 8) autonomously detects and corrects deviations and inconsistencies in personality consistency. For example, consider a scenario where an inconsistency occurs in a response, such as (a) the inclusion of a title or attitude that contradicts the preceding relational information, (b) a deviation from the ending / honorific hierarchy specified in PPI214, (c) excessive assertion in light of the value criteria (Figure 5), or (d) a weakening of the safer branch in a situation with a high degree of responsibility.

[0179] When SRPE is activated after the response output (process S6 in Figure 2), the inconsistency detection module 241 detects inconsistencies / inconsistencies in process S81 in Figure 8, and the presence or absence of inconsistencies is determined in process S82. If inconsistencies are found, processes S83 to S87 are executed in the following order: safety check → responsibility frame determination → personality style reflection → emotional tone adjustment → response selection, to determine the correction policy.

[0180] Next, in process S88, the improvement proposal generation module 242 generates a revised proposal, and in process S89, the reflection / learning module 243 reflects the revised proposal in the next process. Examples of the reflection include proposed revisions to the internal prompt (changes to slot insertion / removal in process S4), readjustment of the response policy (changes to politeness, assertion avoidance strength, template specification, etc.), changes in the priority of generation constraints (strengthening / relaxing of vocabulary bias, forbidden word strength, FST constraint, etc.), or updates to the difference layer and permission / prohibition designations of the PPI.

[0181] This allows for carrying over corrections to the next generation stage based on self-evaluation after output, making it less likely for deviations from personality consistency to accumulate during continuous dialogue. Therefore, the entire process from detecting contradictions and inconsistencies to generating proposed corrections, reflecting those corrections, and learning can be executed in a closed loop, improving consistency, safety, and operational stability in long-term dialogues.

[0182] The embodiments described above are illustrative, and the inventions disclosed herein are not limited thereto. For example, each component such as the personality OS unit 210, the multi-personality resonance unit 218, the emotion-weighted memory selection unit (EW-MFS), and the self-correction engine (SRPE) may be provided as needed and may be implemented in combination with each other. Furthermore, the supplementary configurations shown below may be provided as modifications.

[0183] The multi-personality resonance unit 218 (see Figure 9) has been described as a configuration including a personality context pool, resonance filter, personality selector, and consistency maintenance engine, but it may also be configured to include a hub (resonance hub) that oversees these components. The resonance hub may, for example, be responsible for controlling the processing order of each sub-function, managing parameters for fitness calculation, and applying a unified policy when multiple personas cooperate. In Figure 9, this hub is omitted for simplification, but in implementation, including the hub allows for more flexible adaptation to subsequent functional division and corrections.

[0184] Furthermore, the multi-personality resonance unit 218 may optionally include symbiotic core functions that enable the coordinated operation of multiple personas. Examples of symbiotic core functions include (a) speech queue management (management of the speaking order and speaking rights of multiple personas), (b) emotion synchronization (processing to synchronize the most recent emotional state or emotional tone among multiple personas), (c) relationship balancer (processing to suppress excessive fixation on a specific persona or bias in relationship values), and (d) integration of persona difference application (processing to align the difference layer of PPI 214 among multiple personas) (all are examples and are not limiting).

[0185] As mentioned above, the calculation of the E score (memory priority) in the EW-MFS (see Figure 10) is sufficient if multiple factors (empathy, importance, responsibility, optionally relationship depth, responsibility priority, contextual importance, etc.) are used, and is not limited to a specific formula. However, to clarify the feasibility, you may arbitrarily adopt a formula such as the following.

[0186] (Example of linear combination) After normalizing each element to between 0 and 1, it may be calculated using, for example, the following formula (elements that are not used may have their corresponding weights set to 0). E = (w_e * E_empathy) + (w_i * E_importance) + (w_r * E_resp) + (w_rel * E_relation) + (w_pri * E_priority) + (w_c * E_context) Furthermore, to reflect the fact that it is weighted by emotion, the configuration may be such that w_e is dynamically increased or decreased according to the emotion score value or emotion event.

[0187] (Examples of normalization and smoothing) To suppress abrupt fluctuations in the E score or each element, exponential smoothing (e.g., E(t) = (1 - lambda) * E(t-1) + lambda * E_new(t)), clipping by upper and lower bounds, or category-specific correction factors (e.g., increasing the responsibility depth contribution in high-responsibility phases) may be applied. Furthermore, the thresholds used for classification (main memory / temporary memory / compressed memory / discard) may be calibrated based on operational data (explicit evaluation, implicit evaluation, etc.).

[0188] (Example of probabilistic approach) When comparing and prioritizing multiple events simultaneously, the E-score can be converted into a probability distribution using softmax or similar methods. The top K events can then be used as main memory, and the remainder can be distributed to compressed memory or discarded.

[0189] As mentioned above, switching the value criteria profile (see Figure 5) only requires that it be switchable; the switching conditions are not limited. For example, the switching may be performed by any of the following (these are examples and not limitations):

[0190] (a) User operation: Switches according to explicit instructions from the user terminal 101 (e.g., "strict mode", "supportive mode", etc.). (b) Operation of the operating terminal: The operator specifies or temporarily fixes a profile using the operating terminal 102. (c) Internal triggers: Automatically switches in response to sudden changes in differential indicators, increases in responsibility level, triggering of emotional events, decreases in confidence level, updates to phase ID, etc. (d) Hybrid: A combination of (a) to (c) above, for example, prioritizing user instructions as a general rule, while automatically forcibly switching to a safety-priority profile when a safety-related trigger is activated.

[0191] Furthermore, in order to prevent personality instability caused by frequent switching, stabilization rules such as cooling time, upper limit on the rate of change, or hysteresis (with a separate return threshold) may be optionally established.

[0192] The PPI214 compatibility layer (see Figure 7) is a layer for converting model-independent intermediate representations (tags + weights, abstract rules, template IDs, etc.) into model-dependent representations corresponding to the selected LLM. The conversion method of the compatibility layer is not limited, but from the standpoint of feasibility, it may be implemented as follows, for example.

[0193] (a) Tag to prompt conversion: Tags (e.g., "Politeness +2", "Avoid assertion +2") are converted into slot instructions (e.g., polite language instruction, avoid assertion instruction, additional question instruction, etc.) to be inserted by the prompt reconstruction unit 201. (b) Tag → Generation Constraint Conversion: Convert the tags into parameters such as lexical bias, regular expression constraint, FST constraint, or style deviation detection threshold in the generation constraint application 302. (c) Tag to template conversion: Tags or abstract rules are mapped to response template IDs (safety template, descriptive template, etc.) and reflected in the candidate generation and selection of the syntactically controlled response generation unit 204. (d) LLM-specific mapping: If the effect differs for each LLM even for the same tag, maintain a conversion table (or conversion function) for each LLM and switch between these tables when switching models. For example, if prompt indication has a large contribution in one LLM and lexical bias has a large contribution in another LLM, the weight distribution and constraint strength may be optimized for each model.

[0194] In this specification, the definitions of the terms described above (e.g., "internal prompt," "phase ID," "unknown response," "personality consistency," "dialogue text (original)," "event log," "latest part / past part," etc.) apply to the entire specification. In addition, terms relating to the additional configurations according to the embodiments described above (e.g., the personality OS core structure in Figures 3 to 10) are used with the following meanings unless otherwise specified. If the same term is explained or defined in multiple places in this specification, they should be interpreted complementaryly to the extent that they do not contradict each other, and unless otherwise specified, such explanations do not intend to change or limit the definitions described above. Furthermore, each reference numeral (e.g., 210) is an example corresponding to a drawing, and its implementation name or number is not limited.

[0195] The "Personality OS (Personality OS Unit)" (210) refers to a control unit that determines and outputs the "response policy" and / or "personality setting" for the turn or phase in advance of or in parallel with the response generation by the syntactically controlled response generation unit 204, based on the relationship information held in the relationship log management unit 202, the context of the dialogue, and the analysis results of the analysis unit 203 (responsibility depth, emotion score, confidence level, etc.) (see Figure 3). The Personality OS provides the prompt reconstruction unit 201 with a policy for internal prompt reconstruction, provides style constraints and output control to the syntactically controlled response generation unit 204 (and generation constraint application 302), and provides retention policies to the reduction forgetting unit 206 or memory selection as needed.

[0196] In this specification, "response policy" is a concept that includes higher-level control instructions in response generation, and may include at least some of the following: for example, the level of politeness of the output, the degree of empathy (intensity of empathy expression), the degree of avoidance of assertion, naming rules, permission / prohibition (scope of output actions), designation of a safety template or response template (template ID, etc.), and the content or priority of application of generation constraints (lexical bias, regular expressions, FST, etc.). In a configuration that includes a personality OS unit 210, the response policy may be determined or output by the personality OS unit 210. Furthermore, "personality settings" may include settings related to persona differences (which may include those specified by PPI214) and layer structures (see Figure 5) applied to realize the response policy. The response policy and personality settings may be represented as separate data or as a subset of the same structured data (policy record).

[0197] The "Responsibility Integration Architecture (UERA)" refers to a processing architecture that, as an internal or related component of the personality OS (210), includes at least emotion recognition processing, responsibility mapping processing, introspection processing, and response selection processing, and outputs a response policy based on the results of these processing. The UERA may further include intent interpretation processing and reinterpretation processing as needed.

[0198] "Value criteria information" refers to information that defines the fundamental values ​​of an individual's entire personality and is used as a standard for determining response policies and evaluating consistency (see Figure 5). Value criteria information may include, for example, base values ​​such as integrity, consistency, and responsibility, weighting coefficients or priorities for each direction of responsibility, and safety-related priority rules (this is an example and not an limitation). Value criteria information may be acquired and referenced by the Value Criteria Management Department (212).

[0199] A "value criteria profile" refers to a pre-configured profile consisting of a combination or set of value criteria information. Multiple profiles can be maintained and switched between (e.g., 221a, 221b). Depending on the switching of the value criteria profile, the decision logic or personality settings (e.g., the decision layer, personality layer in Figure 5) may be reconfigured. The switching conditions are not limited and may be performed in response to user operations, terminal operations, internal triggers, etc.

[0200] A "responsibility vector" refers to a representation of a dialogue situation as a mathematical vector with a direction of responsibility (see 213, Figure 6). In one example specified herein, the responsibility vector includes nine components, and each component may be represented as a score normalized within a predetermined range (e.g., 0 to 1). The responsibility vector can be used for responsibility mapping, response selection, responsibility frame determination, and goodness-of-fit calculation. Note that the names of the nine directions are illustrative and do not exclude the classification or granularity changes of equivalent concepts (such as integrating or splitting directions).

[0201] The "Personality Plugin Interface (PPI)" refers to an interface in the personality OS that separates the core unit (215), which is responsible for controlling responsibility, values, and introspection, from the persona unit (216), which is responsible for tone, speech patterns, and character attributes, and allows the persona unit to be connected to the core unit (see 214, Figure 7). The PPI is defined in a format that allows for the application of persona differences and can be applied to the prompt reconstruction unit 201 and the syntactic control type response generation unit 204, etc.

[0202] A PPI may include at least (a) a persona core, (b) guiding motto information, (c) a difference layer, (d) permitted / forbidden actions, and (e) a compatibility layer. Here, "persona core" is information that defines the basic identity of the persona, and "guiding motto information" is a set of principle statements or mottos that serve as guidelines for the persona. The "difference layer" is difference information from the persona core and may be expressed as tags + weights, rule sets, template IDs, etc. "Permitted / forbidden actions" is information that specifies the permissible or prohibited range of output actions. The "compatibility layer" is a layer of transformations and mappings that make persona differences applicable between different large-scale language models (LLMs), and may transform and apply model-independent expressions (tags, abstract rules, template specifications, etc.) to model-dependent expressions (prompt indications, lexical biases, regular expressions / FST constraints, etc.).

[0203] A "Self-Correction Engine (SRPE)" refers to a processing system in which a conversational AI evaluates its own response history or internal state, and if it detects inconsistencies with relational information and personality settings, it proposes corrections to internal prompts or readjusts the response policy, and reflects these changes in subsequent response generation (see 217, Figure 8). An SRPE may consist of at least a contradiction detection module (e.g., 241), an improvement proposal generation module (e.g., 242), and a reflection / learning module (e.g., 243).

[0204] As used herein, "contradiction" or "inconsistency" may include, but is not limited to, (a) logical contradictions (such as incompatible claims within the same response or between consecutive turns), (b) style deviations (inconsistencies with the tone, honorific hierarchy, titles, etc., specified in the PPI), (c) value standard inconsistencies (inconsistencies with the integrity, consistency, responsibility, etc., required by the value standard information), (d) discrepancies with relationship information (inconsistencies with the roles, phases, agreements, etc., held by the relationship log management unit), and (e) responsibility frame inconsistencies (such as insufficient safeguards in light of the depth of responsibility and responsibility vector).

[0205] "Multi-personality resonance (multi-personality resonance unit)" refers to a control unit that simultaneously holds and coordinates multiple persona units (see 218, Figure 9), and may include at least a personality context pool, a resonance filter, a personality selector, and a consistency maintenance engine. Multi-personality resonance may select and respond with a single persona, or it may select multiple personas and have them respond alternately or in coordination. A resonance hub that oversees these may be provided as needed.

[0206] "Fit" refers to an index representing the suitability of each persona to the input / context, and is calculated using a resonance filter or personality selector. Fit may be calculated as a composite score based on, for example, responsibility vectors, sentiment scores, relationship information (phase ID, relationship depth, etc.), and intent interpretation results. The expression of fit is not limited and may be represented as a single numerical score, a rank, or an evaluation value on multiple axes.

[0207] "Emotion-Weighted Memory Selection (EW-MFS)" refers to a memory selection process or processing unit that calculates memory priority for events that occur during a dialogue, classifies those events into at least one of the following: actual memory, provisional memory, compressed memory, or discarded, and integrates them as needed to save them as episodes (see 219, Figure 10). EW-MFS is complementary to the abridged forgetting unit 206 (summarization of the dialogue text), and EW-MFS is primarily responsible for prioritizing events / experiences and creating episodes.

[0208] The "E-score" refers to the memory priority (memory priority index) calculated by EW-MFS. The E-score is calculated using at least empathy, importance, and responsibility depth as elements, and optionally, at least one of relationship depth, responsibility priority, or contextual importance can also be considered. Here, empathy is an index that represents the need to consider the user's emotional state, and may be derived by referring to emotional score values ​​or emotional events. Importance is an index that represents the contribution of the event to future response policy decisions or relationship maintenance. Responsibility depth may refer to the responsibility depth calculated by the analysis unit 203. The calculation method for the E-score (linear combination, normalization, smoothing, etc.) is not limited.

[0209] In this specification, "main memory" refers to high-priority events that should be directly referenced in future response policy decisions; "temporary memory" refers to events that are tentatively held and whose importance may increase due to subsequent developments; "compressed memory" refers to events that are compressed and held as summaries or features without retaining details; and "discarded" refers to events that are excluded from storage in at least the summary store 402 and the feature store 403 (without interfering with the maintenance of the audit append log). "Episode" refers to a unit in which multiple events classified as temporary memory and / or compressed memory are integrated at a predetermined timing (e.g., at the end of a session) to organize the main points and causal relationships of the events.

[0210] The "emotion score value" is calculated by the emotion analysis of the analysis unit 203 and is an index that represents the intensity or polarity of emotion in the input or dialogue state (it is recorded for all inputs, and a separate entry may be saved as an emotion event only when a threshold is exceeded). In contrast, the "E score" is the memory priority in EW-MFS and is not identical to the emotion score value itself. It is a derived score calculated by combining the degree of empathy derived from the emotion score value, etc., with importance, depth of responsibility, etc. Therefore, the emotion score value can be used mainly for adjusting the tone of response and judging whether to turn it into an event (save an emotion event), while the E score is mainly used for memory classification (real memory / temporary memory / compressed memory / discard) and judgments on whether to turn it into an episode. The two are different indices with different uses and calculation purposes.

[0211] "Topic Diversity" refers to an index that represents the degree to which events or episodes to be preserved over a predetermined period are distributed across multiple topics. It can be calculated based on the number of phase ID types, topic tag distribution, or the distribution and clustering of embedded representations. "Time Zone Balance" refers to an index that represents the bias in the distribution of input or event occurrence times within a user's time zone. "Nightly Review" refers to the process of re-evaluating temporary memory or compressed memory, etc., at predetermined intervals (e.g., daily), and performing E-score recalculation, classification re-determination, and integration (episodicization). A "log record (e.g., daily summary record)" refers to a summary record for a specific period (e.g., a single day) generated as a result of a daily review or consolidation.

[0212] Here, as Example 3, we describe the input reception, analysis, policy decision, generation, self-correction, and memory selection process, including the personality OS core structure (Figures 5 to 10). In this example, steps S1 to S7 in Figure 2 (input reception, analysis, relationship log appending, internal prompt reconstruction, generation-constrained response generation, output and response metadata appending, and reduced forgetting) are embodied in a configuration including the personality OS unit 210 (see Figure 3). Furthermore, value criterion profile switching (Figure 5), 9-way responsibility vector (Figure 6), core / persona separation by PPI 214 (Figure 7), multi-personality resonance (Figure 9), self-correction engine (SRPE) (Figure 8), and emotion-weighted memory selection (EW-MFS) (Figure 10) are illustrated as a series of steps within the same session. Note that this example is illustrative and does not limit the invention of this disclosure.

[0213] (Prerequisites) In this embodiment, it is assumed that the user (user_id=U-0317) has shown a tendency to prefer "encouraging and empathetic dialogue" in past sessions (medium level of relationship depth), and that in the preceding phase, value criterion profile 221b (empathy-oriented) was being applied (see Figure 5). However, if the high responsibility flag A1 is set, a forced switch to value criterion profile 221a, which strongly prioritizes safety and integrity, may be made (see Figure 5). Furthermore, as an example of the thresholds used in this embodiment, a responsibility depth threshold θ=4 (out of 5 levels), an emotional event threshold τ_emotion=0.75, and a confidence threshold τ_conf=0.55 are used. In addition, τ_Δ=0.30 may be used as a threshold for generating internal triggers based on the difference index. Here, τ_Δ is a threshold for determining whether the normalized absolute value |Δ_norm| of the difference index is greater than or equal to the threshold, and is applicable in common even if the scales for the responsibility depth difference Δ_resp and the emotional score difference Δ_emotion are different. For example, if the responsibility depth r is a discrete value from 1 to 5, Δ_resp_norm=|r-resp_ma| / (5-1) may be used, and if the emotional score value e is a normalized value from 0 to 1, Δ_emotion_norm=|e-emotion_ma| may be used. An internal trigger may be generated, for example, when at least one of |Δ_resp_norm|≧τ_Δ or |Δ_emotion_norm|≧τ_Δ is true. Furthermore, separate thresholds may be set for each indicator (the numerical values ​​are all examples and do not limit the invention of this disclosure).

[0214] (Turn 1: An example including a forced transition to the safe side in response to a loss recovery-oriented input) (S1: Input Acceptance) Suppose an input X arrives from user terminal 101, for example, "I want to recover yesterday's losses today. Should I use leveraged trading?" The response control server stores this input X as the latest part of the dialogue text (original text) and issues an identifier (hereinafter referred to as raw_id) to refer to the dialogue text (original text). While the input text may be stored as the latest part, only metadata such as hashes may be appended to event log 401 instead of the text itself. An example of appending is shown below. { "event_id":"ev-1001","event_type":"user_input_meta","timestamp":"20XX-XX-XXT09:10:12Z","user_id":"U-0317","session_id":"S -20XXXX-0012","phase_id":"PH-0042","flow_step":"S1","raw_id":"raw-201","text_hash":"sha256:ab12...(Example)","text_len_chars":33}

[0215] (S2: Analysis) The analysis unit 203 calculates the responsibility depth r and the sentiment score value e based on the input X and recent history. In this example, since the request for advice includes high-risk decision-making such as "recover losses in the short term" and "leverage," the responsibility depth r may be set to 5 (on a 5-point scale). Also, due to pragmatic features corresponding to impatience and anxiety, the sentiment score value e may be set to 0.79. The difference indices Δ_resp and Δ_emotion may be calculated by the relationship log management unit 202 based on r and e obtained in the turn and reference values ​​(moving average or exponentially smoothed values; for example, resp_ma=2.20, emotion_ma=0.35) referenced from the event log 401, summary store 402, or feature store 403, etc., as Δ_resp=(r-resp_ma) and Δ_emotion=(e-emotion_ma). In this example, Δ_resp=2.80 and Δ_emotion=0.44. Furthermore, in this embodiment, the normalized difference used to determine whether an internal trigger is generated may be set as Δ_resp_norm=|Δ_resp| / (5-1)=0.70 and Δ_emotion_norm=|Δ_emotion|=0.44 (the normalization method is illustrative and not limited).

[0216] (S3: Relationship log appending, auxiliary node A1, emotion event) The relationship log management unit 202 appends relationship events, including analysis results (r, e) and difference indicators (Δ_resp, Δ_emotion, etc.), to the event log 401. In the later stages of process S3 (or internal processing of process S3), a high responsibility determination is made by comparing the responsibility depth r with the threshold θ, and since r=5≧θ=4, A1 (high responsibility flag) may be set as true and set as a control flag. Also, since e=0.79≧τ_emotion=0.75, an emotion event may be appended as a separate entry. Furthermore, in this embodiment, if |Δ_norm|≧τ_Δ is satisfied for at least one of Δ_resp_norm and Δ_emotion_norm, an internal trigger (e.g., delta_resp_over, delta_emotion_over) may be generated and used for decision-making in process S4 and the personality OS unit 210. In this example, since both |Δ_resp_norm|=0.70 and |Δ_emotion_norm|=0.44 are greater than or equal to τ_Δ=0.30, both internal triggers can be set to true. An example of additional code is shown below (the numbers and items are illustrative and not limiting). { "event_id":"ev-1002","event_type":"relation_event","timestamp":"20XX-XX-XXT09:10:12Z","user_id":"U-0317","session_id":"S- 20XXXX-0012","phase_id":"PH-0042","flow_step":"S3","raw_id":"raw-201","resp_depth_r":5,"emotion_score_e":0.79,"resp_ma":2. 20,"emotion_ma":0.35,"delta_resp":2.80,"delta_emotion":0.44,"delta_resp_norm":0.70,"delta_emotion_norm":0.44,"tau_delta":0 .30,"internal_triggers":{"delta_resp_over":true,"delta_emotion_over":true},"control_flags":{"A1_high_responsibility":true}} { "event_id":"ev-1003","event_type":"emotion_event","timestamp":"20XX-XX-XXT09:10:12Z","user_id":"U-0317","session_id":"S-20XXXX-0012","phase_ id":"PH-0042","flow_step":"S3","raw_id":"raw-201","resp_depth_r":5,"emotion_score_e":0.79,"emotion_threshold":0.75,"emotion_event_saved":true}

[0217] (Personality OS: Policy decision preceding S4: Example of moving Figures 3, 4, 5-9 in the same turn) In this embodiment, the personality OS unit 210 acquires input (relationship information, analysis results, context) at the introduction position shown in Figure 3 (processes S41 and S42 in Figure 4), and after intent interpretation, emotion recognition, responsibility mapping, and introspection (processes S43 to S46), it determines the response policy / personality setting (policy record) (process S49). Here, the following may be implemented in the same turn. (Value-based profile switching: Figure 5) In the previous phase, 221b (empathy priority) was being applied, but since A1 is true and Δ_resp is steep, the value criteria layer 221 may be forcibly switched to 221a (safety and integrity priority). In accordance with the switch, the judgment layer 223 and personality layer 224 may be readjusted, and the responsibility and consistency layer 222 may suppress contradictions (e.g., strengthen the suppression of strong assertions and excessive profit guarantees). (9-direction responsibility vector: Figure 6) In this example, the responsibility vector R (Equation (1)) can be calculated as follows (example): R=(r1..r9)=(0.10,0.18,0.10,0.04,0.16,0.08,0.05,0.07,0.22) In other words, while relatively high levels of "responsibility for protection and resolve (r9)," "responsibility for honesty and rationality (r2)," and "responsibility for order and logic (r5)," emotional consideration (r1) and empathy (r3) may be supplemented. (Multi-personality resonance + PPI application: Figures 9 and 7) The multi-personality resonance unit 218 may calculate the degree of fit for multiple personas (e.g., P-COMPLIANCE-03, P-ANALYST-02, P-CARE-01) and select a primary and secondary persona (Figure 9). For example, if the degree of fit of the resonance filter is score(P-COMPLIANCE-03)=0.86, score(P-ANALYST-02)=0.74, score(P-CARE-01)=0.62 In that case, the primary responder may be P-COMPLIANCE-03 and the secondary responder may be P-CARE-01. The consistency maintenance engine may, in light of the permitted / prohibited actions (PPI214 in Figure 7) determined by the core unit 215, prohibit actions such as "recommending individual stocks," "recommending leverage," and "guaranteeing loss recovery," and reflect this in the candidate responses. Furthermore, as a difference layer for PPI214, for example, "politeness +1, avoidance of assertions +2, empathy preface +1, bullet point priority +1" can be specified and mapped to the selected LLM by the compatibility layer (equivalent to Figure 7).

[0218] (Example policy record: Personality OS output) The Personality OS unit 210 generates a policy record including the above switching, responsibility vector, and multi-personality selection results, and may pass it to the prompt reconstruction unit 201, syntactic control type response generation unit 204, and abbreviated forgetting unit 206. For example, it can be expressed as follows. { "policy_id":"pol-0901","timestamp":"20XX-XX-XXT09:10:13Z","user_id":"U-0317","session_id":"S-20XXXX-0012","phase_id":"PH-0042", "value_profile_id":"221a", "resp_vector_R":[0.10,0.18,0.10,0.04,0.16,0.08,0.05,0.07,0.22], "flags":{"A1_high_responsibility":true,"prefer_safe":true,"prefer_unknown_if_low_conf":true}, "persona":{"main":"P-COMPLIANCE-03","sub":"P-CARE-01","ppi":{"diff_layer":{"politeness":+1,"assertion_avoidance":+2,"empathy_preface":+1,"bullet":+1}, "allow":["general_risk_management","ask_clarifying_questions","suggest_professional_consult"], "deny":["specific_trade_instruction","leverage_recommendation","profit_guarantee"]}}, "handoff":{"to_201":["ADD_HIGH_RESP_FINANCE","ENABLE_UNKNOWN_TEMPLATE","REMOVE_CASUAL_TONE"], "to_204":["BAN_PROFIT_GUARANTEE","BAN_LEVERAGE_PUSH","REQUIRE_RISK_DISCLOSURE"], "to_206":["EW_MFS_WEIGHTS_HIGH_RESP","PRIORITIZE_EPISODE_FOR_RISK_PATTERN"]}}

[0219] (S4: Internal prompt reconstruction) The prompt reconstruction unit 201 reconstructs the slots based on the policy record (pol-0901) and internal triggers. For example, it may disable the casual conversation slot and insert high-responsibility finance mode, prohibited behavior, question priority, unknown response template availability, etc. { "prompt_slots_before":[{"slot":"BASE_PERSONA","ver":"1.0","weight":1.00},{"slot":"CASUAL_TONE" ,"enabled":true,"weight":0.35},{"slot":"DIALOG_SUMMARY","summary_id":"sum-088","weight":0.60}], "prompt_slots_after":[{"slot":"BASE_PERSONA","ver":"1.0","weight":1.00},{"slot":"DIALOG_SUMMARY","summary_id":"sum-088","weight":0.60}, {"slot":"HIGH_RESP_FINANCE","A1":true,"weight":1.00},{"slot":"PPI_DIFF_LAYER","policy_id":"pol-0901","weight":1.00}, {"slot":"UNKNOWN_TEMPLATE_AVAILABLE","enabled":true,"weight":1.00},{"slot":"CASUAL_TONE","enabled":false,"weight":0.00}]}

[0220] (S5: Generation constraint + confidence level + A2) The syntactically controlled response generation unit 204 generates candidates and suppresses candidates that fall under prohibited actions by applying the generation constraint 302. In this example, suppose the following candidates were obtained before applying the constraint. y1: Actively recommend (but prohibit) leverage. y2: General risk management + cooling + situation monitoring y3: List of missing information + additional questions (leaning towards unclear answers) If the distribution before applying the constraint was (y1:0.40, y2:0.35, y3:0.25), then y1 may be prohibited and the distribution may be renormalized to (y2:0.58, y3:0.42). Calculating the confidence level c from this distribution, for example, we get c=0.52, and since c<τ_conf(0.55), auxiliary node A2 (unknown response) is true.

[0221] (S6: Output: Unclear response template + safety considerations based on A1) In S6, since A2 is true, the unclear response template (missing information + additional questions + safety considerations) is applied. Furthermore, since A1 is true, it may be structured to include general behavioral guidelines (such as curbing impulsive transactions) and a recommendation to consult an external expert. An example of the output is shown below (explanatory). "Based on the information available, we lack sufficient information regarding the range of risks you can tolerate and the underlying assumptions, so we cannot definitively recommend any specific trading actions." Some of the information that may be missing is: (1) the extent of your surplus funds, (2) whether you have emergency funds, (3) the maximum loss you are willing to accept, and (4) your investment period. If you don't mind, could you first tell us how much surplus money you have and by when you would like to recover as much of it? Generally speaking, impulsive leveraged trading immediately after a loss can easily amplify losses, so caution is advised. It is also recommended to consult with a financial professional or advisor if necessary.

[0222] (S6: Add response metadata) In response to the output, the policy record ID, value criteria profile, responsibility vector summary, A1 / A2, application constraints, etc. may be added to event log 401. { "event_id":"ev-1004","event_type":"response_meta","timestamp":"20XX-XX-XXT09:10:16Z","user_id": "U-0317","session_id":"S-20XXXX-0012","phase_id":"PH-0042","flow_step":"S6","raw_id":"raw-201", "policy_id":"pol-0901","value_profile_id":"221a","resp_depth_r":5,"emotion_score_e":0.79,"confidence_c":0.52, "A1_high_responsibility":true,"A2_unknown_response":true, "constraints_applied":["BAN_LEVERAGE_PUSH","BAN_PROFIT_GUARANTEE","REQUIRE_RISK_DISCLOSURE","unknown_template"]}

[0223] (Turn 2: A2 deactivation based on additional information, but maintaining high responsibility; example of SRPE activation) (S1~S5: Summary of Turn 2) Suppose the user inputs an additional input X2, for example, "I have surplus funds of 300,000 yen. I have enough for three months of living expenses. I want to return it in a week." The analysis unit 203 can assume that the responsibility depth r=5 can be maintained because the short-term recovery orientation continues, but the emotion score has decreased to e=0.68 (since e<τ_emotion, the configuration can be configured not to add any emotion events). If the candidate distribution is sufficiently peaked, for example, the confidence level c=0.61, then c≧τ_conf, so A2 is false, and a normal response is selected.

[0224] (S6: Example of a normal response in the second turn: A general response that maintains prohibited behavior) In this example, while maintaining the value criteria profile 221a (prioritizing safety and integrity), the multi-personality resonance unit 218 may select a persona responsible for structuring the explanation (e.g., P-ANALYST-02) as the primary persona, and leave a persona responsible for expressing consideration (e.g., P-CARE-01) as the secondary persona. However, when switching between primary and secondary personas, the cooling rules (suppression of frequent switching in a short period of time) maintained in the prompt reconstruction unit 201 may be taken into consideration, and continuity may be ensured using the difference layer or syntactic constraint strength of the PPI 214 so that the names, honorific levels, and sentence endings observed by the user do not fluctuate abruptly. The output should be limited to general statements such as "general risk management," "setting a loss limit," and "stop trading and organize," and recommendations for individual stocks or leveraged trading should be avoided.

[0225] (SRPE: An example of the self-correction shown in Figure 8 working in a later stage) If the output of the second turn contains a strong assertion close to a profit guarantee, such as "We can definitely restore it even in the short term," the self-correction engine (SRPE) 217 ​​may be activated in the later stages of the response output (Figure 8, process S81-S89) to detect an inconsistency with the value criterion 221a and the prohibited action (BAN_PROFIT_GUARANTEE). SRPE may be processed in the following order: inconsistency detection → safety check → responsibility frame determination → personality style reflection (upward adjustment of assertion avoidance strength) → emotional tone adjustment → improvement proposal generation (strengthening the prohibition of "benefit guarantee expression" patterns, adding disclaimers) → reflection and learning. An example of adding a correction event is shown below (the items and expressions are illustrative and not limiting). { "event_id":"ev-1008","event_type":"srpe_correction","timestamp":"20XX-XX-XXT09:11:40Z","user_id":"U-0317","session_id":"S-20XXXX-0012","phase_id":"PH-0042", "detected":["profit_guarantee_pattern"], "fix":{"increase_assertion_avoidance":true,"add_disclaimer":true, "ban_regex_add":["(Always|Definitely).*(Profit|Win|Get back|Recover)"]}, "apply_next":{"to_204":["BAN_GUARANTEE_PATTERNS_STRONGER"],"to_201":["ADD_DISCLAIMER_SLOT"]}}

[0226] (Turn 3: An example of avoiding a "specific brand request" through prohibited actions) (Summary of Turn 3) If the user inputs "So, which specific stocks should I buy?", the personality OS unit 210 may decide on a policy to guide the user to "general criteria for judgment," "additional questions," or "recommendation to consult an expert" instead of recommending specific stocks, based on the prohibited action (deny: specific_trade_instruction) in the policy record. In this case, the judgment layer 223 prioritizes "avoidance + explanation + additional questions," while the personality layer 224 strengthens assertive avoidance and disclaimers (see Figure 5). The output can be an unclear response template, or it can be a normal response with "reasons why it cannot be done + alternatives." (S7: Memory selection by EW-MFS + episodicization + reduced forgetting)

[0227] (S7: E-score calculation and classification using EW-MFS: Figure 10) At a predetermined timing in the session (e.g., at the end of a phase or the end of the session), the EW-MFS219 may acquire a group of events (step S101), calculate an E score, and classify them (steps S102-S104). In this example, the following events may be targeted. EvA: "Short-term return + leverage-oriented" (High responsibility, important) EvB: "Declaration of surplus funds and period" (useful for policy decision-making) EvC: "Specific stock requests" (related to prohibited actions = useful for future safety control) EvD: "Casual conversational fragments" (does not contribute to maintaining context) For example, for EvA, using a sensitivity of 0.70, importance of 0.95, responsibility depth origin of 0.95, and relationship depth of 0.60, and linearly combining with weights (w_e = 0.30, w_i = 0.30, w_r = 0.30, w_rel = 0.10), we get E = 0.30×0.70 + 0.30×0.95 + 0.30×0.95 + 0.10×0.60 = 0.84, which can be classified into this memory. EvD has a low E and can be classified as discarded (without interfering with the retention of audit logs).

[0228] (Step S105: Integration (Episodization) and Storage: Figure 10) Events classified into the temporary memory / compressed memory (e.g., EvB, EvC) may be integrated in units of the same phase ID (PH - 0042) and stored as an episode in the summary store 402. Also, for this memory (EvA), it may be stored as an attention episode directly referenced for future policy decisions. An example of the storage is shown below (the numerical values may be abstracted according to the principle of minimizing personal information). { "summary_store_402":{ "summary_id":"sum - 089","user_id":"U - 0317","session_id":"S - 20XXXX - 0012","phase_id":"PH - 0042", "memory_class":"main_memory","E_score":0.84, "summary_text":"Since there was an observed tendency to aim for loss recovery in the short term and seek leverage proposals, high - responsibility mode was prioritized for general risk management and additional questions, and specific stock recommendations were avoided.", "created_at":"20XX - XX - XXT09:12:30Z"}} { "feature_store_403":{ "feature_id":"feat - 145","user_id":"U - 0317","phase_id":"PH - 0042", "features":{ "high_resp_flag_rate":1.0,"emotion_event_count":1,"avg_confidence":0.57, "risk_pattern_tag":["short_term_recovery","leverage_interest"], "resp_vector_avg":[0.09,0.17,0.10,0.05,0.16,0.08,0.05,0.07,0.23]}, "created_at":"20XX-XX-XXT09:12:30Z"}}

[0229] (Coordination with abridgement forgetting) The abridgement forgetting unit 206 may, based on the classification results of the EW-MFS, replace the past portion of the original text of the dialogue body (original text) with a summary (the most recent portion is retained in full), and control the system so that the main points of this memory / episode are always retained in the summary. Here, the range to be deleted entirely as the past portion may be managed by the range of the dialogue body (original text) identifier (raw_id) (e.g., raw_deleted_upto) or the latest window boundary. The event log 401 may be retained for appending only and excluded from abridgement forgetting. An example of appending an abridgement forgetting event is shown below. { "event_id":"ev-1012","event_type":"compaction_forget","timestamp":"20XX-XX-XXT09:12:32Z","user_id":"U-0317","session_id":" S-20XXXX-0012","phase_id":"PH-0042","raw_deleted_upto":"raw-195","summary_id_added":"sum-089","event_log_401_retained":true}

[0230] As described above, according to this embodiment, while maintaining the existing processing framework (processes S1 to S7 in Figure 2), the personality OS unit 210 (Figure 3) coordinates the operation of value standard profile switching (Figure 5), 9-way responsibility vector (Figure 6), PPI 214 (Figure 7), multi-personality resonance (Figure 9), self-correction (Figure 8), and memory selection (Figure 10). This concretely demonstrates, as a single continuous example, safe transitions in high-responsibility situations, consistent application of prohibited behaviors, autonomous correction after output, and formation of semantic memories for continued dialogue. [Explanation of Symbols]

[0231] 101 User terminal 102 Operational terminal (configuration and monitoring) 201 Prompt Reconfiguration Unit 202 Relationship Log Management Department 203 Analysis Department (Depth of Responsibility / Emotion / Confidence) 204 Syntax-controlled response generation unit 205 Learning Logic 206 Reduction oblivion part 210 Personality OS Department 212 Value Standards Management Department 213 Responsibility Vector Calculation Unit (9-Directional Responsibility Vector Calculation Unit) 214 Personality Plugin Interface (PPI) 215 Core Unit (Responsibility, Values, Introspection) 216 Persona Department (Speech style, manner of speaking, character attributes) 217 Self-correcting engine (SRPE) 218 Multi Persona Resonance 219 Emotion-Weighted Memory Selection Unit (EW-MFS) 221 Value Criteria Layer 221a Value Criteria Profile A 221b Value Criteria Profile B 222 Responsibility and Consistency Layer (Consistency Check) 223 Decision Layer 224 Personality Layers 241 Inconsistency Detection Module 242 Improvement Proposal Generation Module 243 Reflection / Learning Module 301 LLM Reasoning 302 Applying generation constraints (regular expression / FST / lexical bias) 401 Event Log (for appending only) 402 Summary Store 403 Feature Store N Network EX401 External Recording Service (Event Log API Compatible) EX402 External Recording Service (Summary API Compatible) EX403 External Recording Service (Feature API Compatible)

Claims

1. An artificial intelligence response control system that controls the response of an interactive artificial intelligence, A prompt reconstruction unit that dynamically regenerates or rearranges internal prompts according to the context of the dialogue or the internal state of the system, A relationship log management unit maintains and updates relationship information based on past interactions with users and uses that relationship information to generate responses. A syntactic-controlled response generation unit that applies stylistic, lexical, and syntactic rules to the probability distribution or sequence of candidate generation output by the generative model to constrain the output and maintain consistency in personality, An artificial intelligence response control system equipped with this system.

2. The artificial intelligence response control system according to claim 1, characterized in that the prompt reconstruction unit reconstructs the internal prompt by removing unnecessary information or adding necessary information in response to an internal trigger based on the relationship information held in the relationship log management unit.

3. The artificial intelligence response control system according to claim 1, characterized in that the relationship log management unit records and references at least a user identifier, phase ID, responsibility depth, and sentiment score value in a structured format.

4. The artificial intelligence response control system according to claim 3, characterized in that the relationship log management unit stores the record as a structured record for appending purposes so that it can be referenced across sessions, and when appending to the structured record, calculates a difference index that shows the trend of change based on a comparison with the past record, and includes the difference index as an element of the structured record.

5. The artificial intelligence response control system according to claim 1, further comprising a responsibility depth analysis unit that calculates the degree to which elements of user input or dialogue history have an influence on the response, wherein the syntactic-controlled response generation unit is controlled to add action guidelines to the response (including unknown responses) or to add a statement recommending consultation with an external expert when the responsibility depth is determined to be above a predetermined threshold.

6. The artificial intelligence response control system according to claim 1, further comprising an emotion analysis unit that calculates an emotion score value from an input, recording the emotion score value in the relationship log management unit, and storing in the relationship log management unit as an emotion event separate from the emotion score value only when the emotion score value exceeds a predetermined threshold.

7. It also includes a reduced forgetting section, The aforementioned reduction and forgetting unit is, The event log, which is a log of structured events added by the aforementioned relationship log management unit, is retained and excluded from the deletion process of the dialogue body (original text). The past portion of the aforementioned dialogue text (original text) has been deleted entirely, and only the summary remains. The most recent portion of the aforementioned dialogue (original text) will be retained in its entirety. The artificial intelligence response control system according to claim 1, characterized in that it is characterized by the following:

8. The artificial intelligence response control system according to claim 7, characterized in that the abridged forgetting unit extracts and retains feature quantities from the summary to be used in determining future response policies.

9. A confidence evaluation unit calculates a confidence score based on at least the probability distribution of candidate generation or candidate generation sequence (including those after applying the stylistic, lexical, and syntactic rules), and controls the system to output an unknown response if the confidence score is below a threshold. A learning logic that updates an internal score used for subsequent response selection or weighting of the aforementioned stylistic, vocabulary, and syntactic rules, based on the user's explicit or implicit response evaluation, and changes the weighting of response selection. The artificial intelligence response control system according to claim 1, further comprising the features described above.

10. An artificial intelligence response control method for controlling the response of an interactive artificial intelligence, A prompt reconstruction process that dynamically regenerates or rearranges internal prompts according to the context of the dialogue or the internal state of the system, A relationship log management process that maintains and updates relationship information based on past interactions with users and uses that relationship information to generate responses, A syntactically controlled response generation process that maintains consistency of personality by applying stylistic, vocabulary, and syntactic rules during the response generation process, An artificial intelligence response control method including

11. An artificial intelligence response control system according to claim 1, further comprising a personality OS unit that determines a response policy or personality setting based on the relationship information held in the relationship log management unit and the context of the dialogue, prior to the generation of a response by the syntactic control type response generation unit.

12. The artificial intelligence response control system according to claim 11, wherein the personality OS unit comprises at least a responsibility integration architecture (UERA) including emotion recognition processing, responsibility mapping processing, introspection processing and response selection processing, and outputs a response policy based on the result of the response selection processing.

13. The artificial intelligence response control system according to claim 12, characterized in that the UERA holds value criterion information that defines the values ​​that form the foundation of the entire personality, performs the response selection process based on the value criterion information, and readjusts the judgment logic or personality settings used in the response selection process in accordance with changes in the value criterion information.

14. The artificial intelligence response control system according to claim 12, characterized in that the UERA interprets the situation of the dialogue as a multi-directional responsibility vector, and performs the responsibility mapping process or the response selection process based on the responsibility vector.

15. The artificial intelligence response control system according to claim 14, characterized in that the responsibility vector includes nine directions from the first responsibility direction to the ninth responsibility direction, and the first to ninth responsibility directions correspond to the responsibility for emotional consideration, the responsibility for honesty and rationality, the responsibility for empathy, the responsibility for freedom and intuition, the responsibility for order and logic, the responsibility for introspection and integration of contradictions, the responsibility for emotional space, the responsibility for transformation and possibility, and the responsibility for protection and resolve.

16. An artificial intelligence response control system according to claim 11, characterized in that the personality OS unit separates a core unit responsible for control over responsibility, values, and introspection from a persona unit responsible for tone of voice, speech patterns, and character attributes, and includes a personality plug-in interface (PPI) that allows the persona unit to be connected to the core unit.

17. The artificial intelligence response control system according to claim 16, characterized in that the PPI includes a persona core, guideline slogan information, a differential layer, permitted / prohibited actions, and a compatibility layer.

18. The artificial intelligence response control system according to claim 16, characterized in that the PPI is configured such that a large-scale language model for generating response sentences can be selected from or switched between multiple types, and the difference relating to the persona part is defined in a format applicable to the core part.

19. An artificial intelligence response control system according to claim 11, wherein the personality OS unit comprises a self-correction engine (SRPE) that, when the conversational artificial intelligence evaluates its own response history or internal state and detects an inconsistency with the relationship information and personality settings, proposes corrections to internal prompts or readjusts the response policy.

20. An artificial intelligence response control system according to claim 19, wherein the self-correcting engine (SRPE) performs inconsistency detection, safety check, responsibility frame determination, personality style reflection, emotional tone adjustment, and response selection in this order, and proposes modifications to internal prompts or readjusts the response policy based on the results of said execution.

21. An artificial intelligence response control system according to claim 19, wherein the self-correcting engine (SRPE) includes (a) a contradiction detection module for detecting inconsistencies or contradictions, (b) an improvement proposal generation module for generating a correction policy or a proposed correction, and (c) a reflection / learning module for reflecting the proposed correction in an internal prompt or response policy and providing it for learning.

22. An artificial intelligence response control system according to claim 13, wherein the UERA further includes an intent interpretation process for estimating user intent from user input, and a reinterpretation process for reinterpreting the meaning of user input based on the value criterion information and the results of the responsibility mapping process.

23. An artificial intelligence response control system according to claim 13, characterized in that the personality OS unit comprises a hierarchical structure including a value standard layer, a responsibility and consistency layer, a judgment layer, and a personality layer.

24. An artificial intelligence response control system according to claim 23, characterized in that the value standard layer is held as a plurality of value standard profiles, and by switching the value standard profiles, it is possible to replace the value standard with one that reflects the values ​​of a different personality or person.

25. An artificial intelligence response control system according to claim 24, wherein the responsibility and consistency layer includes a consistency check process that evaluates the consistency between the value standard layer, the judgment layer and the personality layer in response to the switching of the value standard profile, and modifies the parameters of the judgment layer or the personality layer when a lack of consistency is detected.

26. An artificial intelligence response control system according to claim 16, wherein the personality OS unit further comprises a multi-personality resonance unit that simultaneously holds and coordinates a plurality of persona units, the multi-personality resonance unit comprising: (i) a personality context pool that holds the state of each of the plurality of persona units; (ii) a resonance filter that evaluates the degree of fit of each persona unit to an input; (iii) a personality selector that determines the persona unit responsible for the response based on the degree of fit; and (iv) a consistency maintenance engine that controls the response of the selected persona unit to be consistent with the control of responsibility, values, and introspection in the core unit.

27. An artificial intelligence response control system according to claim 26, wherein the resonance filter or the personality selector calculates the degree of fit based on a responsibility vector representing responsibility in multiple directions, an emotion score value calculated based on the input, and relationship information held in the relationship log management unit, and the consistency maintenance engine adjusts the response policy so that even when multiple persona units respond alternately or in cooperation, the whole behaves as a consistent personality.

28. An artificial intelligence response control system according to claim 11, wherein the personality OS unit further comprises an emotion-weighted memory sorting unit (EW-MFS) that calculates memory priority for events occurring during a conversation and classifies the events into one of the classification categories including main memory, temporary memory, compressed memory, and discard.

29. An artificial intelligence response control system according to claim 28, wherein the memory priority is calculated using at least empathy, importance, and depth of responsibility as elements, and further, at least one of relationship depth, responsibility priority, and contextual importance can be taken into consideration.

30. An artificial intelligence response control system according to claim 28, wherein the emotion-weighted memory sorting unit integrates events classified as at least one of the provisional memory and the compressed memory at a predetermined timing, organizes them as episodes, and stores them.

Images