Agent-oriented information management method, device, storage medium and program product

Abstract

Embodiments of the present application provide an agent-oriented information management method, device, storage medium and program product. In the embodiments of the present application, through a multi-dimensional classification label mechanism, the multi-dimensional classification labels of the agent and the historical conversation context of the target object are associated and stored, a multi-level memory organization architecture is constructed, so that in subsequent interactions, the agent can quickly recall information related to the user interaction request from the historical conversation context based on the multi-dimensional classification labels, breaking through the single conversation life cycle, and enabling effective reuse of knowledge in historical conversation content through multi-dimensional classification label matching between different conversations, thereby cross-conversation knowledge reuse, which helps to enhance the long-term memory ability of the agent, and promotes the evolution of the AI system such as the agent from a tool for executing isolated tasks to an intelligent asset carrying organizational experience and wisdom.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an information management method, device, storage medium and program product for intelligent agents. Background Technology

[0002] With the rapid development of Large Language Models (LLM) and intelligent agent technology, artificial intelligence (AI) systems have been widely applied in various interactive scenarios. However, most current AI systems generally adopt a session-level context design in their memory mechanism. After a single session ends, the information generated by that session interaction is cleared or discarded, resulting in a lack of continuity and personalization in AI interaction.

[0003] In enterprise-level applications, the drawbacks of this short-term memory are particularly pronounced. Interactions in an enterprise environment are characterized by their long-term, cumulative, and knowledge-intensive nature. Communication between employees and AI systems is not only about completing immediate tasks, but also a crucial vehicle for the accumulation and reuse of organizational knowledge and experience. The inability of AI systems to effectively accumulate and reuse past information leads to efficiency losses and knowledge gaps, making it difficult for AI systems to evolve from tools for performing isolated tasks into intelligent assets that carry organizational wisdom. Summary of the Invention

[0004] This application provides an information management method, device, storage medium, and program product for intelligent agents, which enables cross-session knowledge reuse of intelligent agents.

[0005] This application provides an information management method for intelligent agents, including: Obtain the historical session context generated from the interaction between the target object and the agent in historical sessions; Based on multi-dimensional classification rules, the historical session context is classified in multiple dimensions to determine the first multi-dimensional classification label of the historical session context. The first multi-dimensional classification label is associated with and stored in the historical session context, so that the agent can retrieve the target information required for the user interaction request from the historical session context based on the first multi-dimensional classification label.

[0006] This application embodiment also provides an information management method for intelligent agents, which associates and stores the historical session context and the multi-dimensional classification label of the historical session context in at least one historical session between the intelligent agent and the target object. The method further includes: Obtain user interaction requests; The user interaction request is subjected to semantic analysis to obtain the interaction intent reflected by the user interaction request; Based on the session identifier corresponding to the user interaction request, determine the user boundary identifier for the current interaction; The user boundary identifier and the interaction intent are mapped to multi-dimensional classification labels to determine the second multi-dimensional classification label corresponding to the user interaction request; From the historical session context of the at least one historical session, the target historical session context associated with the second multi-dimensional classification label is selected; Based on the interaction intent, information retrieval is performed in the target historical session context to obtain the target information required for the user interaction request.

[0007] This application also provides an electronic device, including: a memory and a processor; wherein the memory is used to store computer programs; The processor is coupled to the memory and is used to execute the computer program to perform the steps in the above-described information management methods for intelligent agents.

[0008] This application also provides a computer-readable storage medium storing computer instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps in the above-described information management methods for intelligent agents.

[0009] This application also provides a computer program product, including a computer program that, when executed by one or more processors, causes the one or more processors to perform the steps in the above-described information management methods for intelligent agents.

[0010] In this embodiment, a multi-dimensional classification tagging mechanism is used to store the multi-dimensional classification tags and associations of the historical session context of the agent and the target object, constructing a multi-level memory organization architecture. This enables the agent to quickly recall information related to the user's interaction request from the historical session context based on the multi-dimensional classification tags in subsequent interactions. This breaks through the life cycle of a single session and enables the effective reuse of knowledge in the historical session content through multi-dimensional classification tag matching between different sessions. This cross-session knowledge reuse and context continuation helps to enhance the agent's long-term memory capabilities and promotes the evolution of AI systems such as agents from tools for performing isolated tasks to intelligent assets that carry organizational experience and wisdom. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating the information management method for intelligent agents provided in an embodiment of this application; Figure 2 A flowchart illustrating another information management method for intelligent agents provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the process of information retrieval based on interaction intent, provided as an embodiment of this application. Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. The various neural network models (such as deep learning models, LLM, multimodal models, etc.) involved in the embodiments of this application comply with relevant laws and standards.

[0014] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating an agent-oriented information management method provided in an embodiment of this application. This information management method is primarily used to store the agent's historical session context. For example... Figure 1 As shown, this information management method mainly includes the following steps: 101. Obtain the historical session context generated by the interaction between the target object and the agent in the historical session.

[0016] 102. Based on the multi-dimensional classification rules, perform multi-dimensional classification on the historical session context to determine the first multi-dimensional classification label of the historical session context.

[0017] 103. The first multi-dimensional classification label is associated with and stored in the historical session context so that the agent can retrieve the target information required for the user interaction request from the historical session context based on the first multi-dimensional classification label.

[0018] In this embodiment, an agent refers to a software entity capable of perceiving the environment, making autonomous decisions, and taking actions to achieve a specific goal. The agent is an AI agent, intelligent code based on a deep learning model. It uses the deep learning model as its "core brain," possessing autonomous understanding, planning, memory, and tool usage capabilities, and is an intelligent system capable of interacting with the external environment to achieve a specific goal. The deep learning model can be a neural network model with a relatively large number of parameters. LLM is merely one example; this embodiment does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning model involved in this embodiment can be an artificial intelligence-based language model (LM) or multimodal model (MM), etc.

[0019] The target object refers to the entity that interacts with the intelligent agent and is the source of the historical session context. In this embodiment, the specific form of the target object is not limited; it can be a natural person user, an enterprise tenant, a digital human instance, a device terminal, or any entity capable of initiating an interaction request to the intelligent agent. In this embodiment, the target object is used to define the boundary of the historical session context generation, ensuring that subsequent classification and recall operations are performed on the historical information of a specific interacting entity, rather than on all indiscriminate data.

[0020] A historical session refers to at least one complete interaction between the target object and the agent that occurred before the current user interaction request. The time frame of a historical session can cover completed sessions between the target object and the agent within a set time period. This time period can be one month, two months, one year, or longer than the current time. Alternatively, it can include completed sessions from the first interaction between the target object and the agent to the time before the current user interaction request. A historical session is the unit that carries the historical session context; each historical session corresponds to an independent interaction record, and the context generated from one or more historical sessions can be obtained for subsequent processing.

[0021] Historical conversation context refers to the set of information generated during the interaction between the target object and the agent in historical conversations. It includes the target object's input content, the agent's reasoning process and response content, interaction timestamps, and conversation identifiers, among other raw interaction data. Historical conversation context is an unstructured natural language dialogue record that has not undergone semantic parsing or classification processing. It is the direct processing object of the multi-dimensional classification rules in this embodiment and is also the original data carrier that is ultimately associated, stored, and bound to the multi-dimensional classification labels.

[0022] Since the historical session context generated by the interaction between the target object and the agent in the historical session is the original data foundation that carries past interaction data and supports cross-session knowledge reuse, the historical session context generated by the interaction between the target object and the agent in the historical session can be obtained in step 101.

[0023] In some embodiments, during the historical session interactions between the target object and the agent, the interaction data of each round of interaction can be automatically written to a persistent log file or message queue. When it is necessary to obtain the historical session context, the log parsing module reads the interaction data corresponding to the historical session in batches from the log storage according to the session identifier of the historical session, and aggregates them into the historical session context. This embodiment uses an asynchronous parsing method of session logs to obtain the historical session context, which can complete data retention without intruding on the core interaction link of the agent, avoid the impact of real-time capture on the interaction response latency, and ensure service performance in high-concurrency scenarios.

[0024] Because agents generally employ a session-level context management mechanism, the interaction data generated in a single session is automatically cleared after the session ends, making it impossible for historical session contexts to be obtained and reused in subsequent processes. To overcome this deficiency, in some embodiments, hook functions can be pre-set in the agent, and these pre-set hook functions can capture historical interaction data generated in the historical session in real time during the interaction between the target object and the agent, serving as the historical session context.

[0025] Hook functions can capture historical interaction data in real time during historical sessions. Hook functions can be event monitors, interceptors, callback functions, or any program module capable of non-intrusive data collection during interaction; this application embodiment is not limited to these. Hook functions can be attached to key nodes in the interaction chain of an agent, such as the entry and exit points of message receiving middleware, dialogue state managers, or response generation modules. When the target object initiates a user interaction request or the agent generates response content, the hook function is automatically triggered, synchronously collecting raw data such as user input text, agent output text, interaction timestamps, session identifiers, and role identifiers from the current interaction round, without waiting for the session to end or relying on asynchronous feedback from an external log system. Furthermore, the captured historical interaction data can be aggregated according to the session dimension to obtain the historical session context corresponding to that historical session.

[0026] This embodiment embeds the acquisition of historical session context into the interaction process itself, without relying on asynchronous retrieval or external log feedback after the session ends, thus achieving real-time data retention before clearing the session context. Simultaneously, it ensures that subsequent multi-dimensional classification processing can be performed based on real, continuous original interaction records, guaranteeing that subsequent multi-dimensional classification and associated storage have a real, continuous, and lossless original data foundation. This effectively overcomes the technical obstacle of traditional intelligent agents' "use and forget" nature, which makes historical information unavailable.

[0027] In some embodiments, the dimension of the target object may be the user dimension. Accordingly, the historical sessions between the target object and the agent may be: a personal session between the user identifier corresponding to the target object and the digital human associated with the agent, or a group session between the digital human and the user identifier and digital human in the group to which the target object belongs. Alternatively, the historical sessions may encompass both the aforementioned personal sessions and group sessions.

[0028] In personal conversation scenarios, hook functions can be used to capture historical interaction data (defined as first historical interaction data) between the user identifier and the digital human associated with the agent, within the personal conversation of the target object's user identifier. This data serves as the historical conversation context. Hook functions capture one-to-one interactions between the user identifier and the digital human; the resulting first historical interaction data reflects the user's private dialogue with the digital human and does not include interaction information from other users. This type of data typically carries information such as the user's personalized preferences and private transaction records, and is a core source for building user-level memory.

[0029] In group conversation scenarios, hook functions can be used to capture historical interaction data between a user identifier and a digital human (defined as second historical interaction data) as historical conversation context. Specifically, the hook function captures interactions between a user identifier and the digital human within the same group. The resulting second historical interaction data extracts dialogue snippets between that specific user and the digital human within the group, rather than the interactions of all members of the group. This selective capture preserves the user's interaction trajectory with the digital human within the group context while avoiding noise interference from irrelevant member information, ensuring that subsequent classification and labeling can accurately distinguish between personal private memories and personal memories related to the user within the group.

[0030] In other embodiments, the aforementioned first and second historical interaction data can also be used as historical session context. Using both the first and second historical interaction data as the historical session context can comprehensively cover the entire interaction trajectory of the target object with the digital human at the user level. This preserves private preferences and personalized transaction records in individual sessions while incorporating interaction fragments between the user and the digital human in a group context within group sessions, avoiding memory gaps caused by single session types. This provides a structured data foundation for targeted recall based on the user level. Simultaneously, the unified collection of personal and group session data at the user level supports coherent knowledge reuse by the agent across different session types. For example, preferences confirmed in personal sessions can be automatically adapted and invoked in group collaboration scenarios, helping to overcome the memory fragmentation defects caused by traditional solutions that can only handle single session types.

[0031] In other embodiments, the target object dimension is a combination of user and group dimension, that is, extracting the interaction data of a user in a certain group. Accordingly, historical sessions include: group sessions between the digital human associated with the agent and the user identifier corresponding to the target object in the group, and no longer include personal sessions between the digital human and the target object. In this embodiment, the historical interaction data between the user identifier and the digital human (defined as third historical interaction data) can be captured in the group session through a hook function as the historical session context. This dimension design in this embodiment is suitable for scenarios that require memory management based on the user's identity unit in a specific group, such as enterprise department collaboration groups, project-specific groups, etc., where the user's interaction behavior is highly dependent on the group context, and personal sessions outside the group do not have independent reuse value. When capturing third historical interaction data, the hook function can carry the group identifier as context metadata in addition to the regular interaction content, so that the subsequently generated multi-dimensional classification labels include the binding relationship between users and groups, which can support knowledge retrieval of the user and group combined dimension.

[0032] After obtaining the historical session context, in step 102, the historical session context can be classified in multiple dimensions according to the multi-dimensional classification rules to determine the multi-dimensional classification label of the historical session context.

[0033] Among them, the multi-dimensional classification rules are predefined classification criteria that describe the set of classification dimensions of the session context and the judgment logic for each classification dimension. These multi-dimensional classification rules can be manually configured explicit rule sets or implicit instructions encoded in the form of prompt words or function definitions. They provide a unified and executable semantic parsing framework for subsequent classification of historical session contexts, avoiding arbitrariness and non-reproducibility in the classification process. Step 102, which performs multi-dimensional classification of the historical session context according to the multi-dimensional classification rules to obtain multi-dimensional classification labels for the historical session context, is the core step in transforming unstructured interaction data into an indexable structured architecture.

[0034] The multi-dimensional classification rules encompass multiple classification dimensions. "Multiple" refers to two or more dimensions. Each dimension corresponds to an independent semantic or administrative attribute, collectively forming a complete characterization of the historical conversation context. The classification rules for each dimension are independently determined based on the content features, metadata information, or semantic representation of the historical conversation context. Finally, the classification results from each dimension are combined to form a multi-dimensional classification label. This step assigns multi-dimensional semantic coordinates to discrete, unstructured dialogue records. Subsequent information retrieval based on user interaction requests can achieve local retrieval rather than a full traversal, helping to improve retrieval accuracy and efficiency.

[0035] In some embodiments, the classification dimension includes: the attribution boundary dimension. Correspondingly, the multi-dimensional classification labels include: the attribution boundary label. The attribution boundary refers to the memory isolation and sharing control dimension based on the reusability of the knowledge content itself carried by the historical session context, used to determine the visibility scope and access permissions of that memory. The attribution boundary is based on the reusability of the knowledge content, ensuring that memory management is centered on content value. Reusability refers to whether the knowledge content carried by the historical session context can be effectively accessed by other users, tenants, or digital humans. For example, "Memory Fault Troubleshooting Steps" is reusable, while "My Frequently Used Keyboard Shortcut Settings" is not.

[0036] Accordingly, ownership boundaries include private boundaries and global boundaries. Similarly, ownership boundary labels include private boundary labels and global boundary labels. A private boundary refers to one type of ownership boundary, corresponding to knowledge content that lacks cross-domain (cross-user, cross-tenant, or cross-digital human, etc.) reusability. This type of memory can only be recalled by a specific target object, preventing privacy leaks and data unauthorization. For example, a user's private information and dedicated development environment configurations fall under the category of private boundaries.

[0037] Global boundaries are another type of ownership boundary, corresponding to knowledge content with cross-domain reusability. This type of memory can be recalled in interactions between authorized users and digital humans within the corresponding group or tenant scope, supporting the accumulation and sharing of organizational-level knowledge. For example, internal enterprise expense reimbursement process specifications, product usage guides, team collaboration agreements, or objective domain knowledge all fall under the category of global boundaries. For instance, the discussion of "Python list comprehension syntax" in a personal conversation falls under the global boundary due to its cross-domain reusability; while mentioning "my cat's name is DouDou" in a group chat, due to its lack of reusability, falls under the private boundary.

[0038] Accordingly, the classification rules for the attribution boundary label may include: whether the knowledge content carried by the historical session context has knowledge reusability. If the knowledge content does not have knowledge reusability, the attribution boundary label is a private boundary label; if the knowledge content has knowledge reusability, the attribution boundary label is a global boundary label.

[0039] In some embodiments, when the ownership boundary label is determined to be a private boundary, the classification dimension further includes: the task scenario dimension. Multi-dimensional classification labels include: task scenario labels. A task scenario refers to the interaction context dimension based on the temporal regularity of tasks, under the premise that the ownership boundary is a private boundary. It is used to distinguish the stability and periodicity of the agent's task execution, providing a basis for subsequent differentiated lifecycle management. Its core focus is whether the task occurs repeatedly at a fixed rhythm, rather than the specific content of the task. Based on the temporal regularity of tasks, task scenarios include: steady-state task scenarios and occasional task scenarios. Steady-state tasks refer to task types with high temporal regularity, typically corresponding to routine tasks executed daily by the agent, with fixed usage processes, standardized outputs, and stable semantic expressions. For example, a stock agent generates a daily stock market analysis report, an operations and maintenance agent outputs a weekly server health report, and a customer service agent summarizes user feedback statistics monthly.

[0040] Occasional tasks refer to task types within a task scenario that lack a regular time pattern. They are typically temporary, non-standard, and unconventional transactions, with semantic expressions that are more random and specific. Examples include troubleshooting sudden network failures, preparing for impromptu meetings, verifying specific data, and handling emergency events. Correspondingly, task scenario tags include: steady-state task tags and occasional task tags.

[0041] In other embodiments, when the attribution boundary label is determined to be a private boundary, the classification dimension also includes: a content attribute dimension. Multi-dimensional classification labels include: content attribute labels. Content attributes, under the premise that the attribution boundary is a private boundary, refer to the semantic topic dimension based on the specific semantic division of knowledge content. They are used to accurately describe the essential category of the dialogue content, providing semantic anchors for content matching and retrieval. Their focus is on what this memory is saying.

[0042] Based on semantics, content attributes can include: preference profiles, task execution, factual knowledge, thought processes, and non-task interaction content. Correspondingly, content attribute tags include: preference profile tags, task execution tags, factual knowledge tags, thought process tags, and non-task interaction tags. Among these, preference profiles are one type of content attribute tag, corresponding to knowledge content that reflects the target audience's personalized behavioral norms, habitual tendencies, or subjective preferences. For example, "I'm used to using the Vim editor" or "I prefer a concise reporting style." Vim is a highly customizable text editor.

[0043] Task execution is a content attribute that corresponds to knowledge content reflecting the task execution trajectory, operation records, or process nodes that conform to the semantics of intelligent agent applications. For example, "Code merging completed and submitted for testing" or "Weekly report sent to department group".

[0044] Factual knowledge refers to knowledge content that reflects objective cognition, public rules, or verifiable information, and does not include subjective judgments or personally attached information. Examples include "the principle of the TCP three-way handshake" and "Python 3.10 supports pattern matching syntax."

[0045] The logic of thought reflects the knowledge content of the decision-making derivation chain, problem analysis process, or reasoning steps, embodying the thinking path of the intelligent agent or user. For example, "Because the server CPU is consistently above 90% and the logs show a memory leak, it is determined that the service needs to be restarted and the code needs to be investigated."

[0046] Non-task interaction content refers to casual conversation, emotional expressions, or irrelevant corpora that do not correspond to a specific task of the agent, have no clear application goal, and do not correspond to a specific application goal. For example, "The weather is really nice today" or "I'm a little tired."

[0047] In some embodiments, when the attribution boundary label is determined to be a private boundary, the classification dimension also includes: a storage type dimension. The multi-dimensional classification label includes: a storage type label. Storage type is based on data resource consumption characteristics, and the classification rule corresponding to the storage type label can be implemented as: whether there are resource-intensive files among the data types contained in the historical session context. Resource-intensive files refer to binary data files that consume more computing, storage, or network resources than ordinary text data during loading, transmission, or processing, and may include: videos, images, medical images, AI datasets, and / or 3D models, etc. If there are resource-intensive files among the data types contained in the historical session context, the storage type label of the historical session context is determined to be a compressed storage label; if there are no resource-intensive files among the data types contained in the historical session context, the storage type label of the historical session context is determined to be a plain text storage label, thereby achieving adaptive allocation of storage resources.

[0048] The classification dimensions and corresponding classification labels for classifying historical session contexts shown in the foregoing embodiments are merely illustrative and do not constitute a limitation. The classification results of the four classification dimensions are combined to form a complete first multi-dimensional classification label, which can be stored in structured key-value pairs associated with the historical session context.

[0049] By classifying historical conversation contexts in multiple dimensions, discrete and unstructured dialogue records can be given multi-dimensional semantic tags. Subsequently, targeted retrieval based on multi-dimensional classification tags can be achieved instead of full traversal, which can improve recall accuracy and efficiency. Furthermore, memory isolation is achieved based on the ownership boundary dimension, reducing the risk of privacy leakage and data overreach. At the same time, it provides a standardized index foundation for cross-conversation knowledge reuse.

[0050] In some embodiments, the historical session context and multi-dimensional classification rules can be jointly input into the agent-dependent deep learning model. For a description of the implementation of the agent-dependent deep learning model, please refer to the relevant content in the foregoing embodiments, which will not be repeated here. Furthermore, the deep learning model can be used to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules to obtain the multi-dimensional classification label of the historical session context. In this embodiment, the agent-dependent deep learning model performs end-to-end multi-dimensional classification of the historical session context according to the multi-dimensional classification rules, directly outputting the multi-dimensional classification label of the historical session context. This method utilizes the semantic understanding and instruction-following capabilities of deep language models, enabling accurate parsing of implicit semantic features and achieving precise classification.

[0051] Specifically, for attribution boundary labels, deep learning models can be used to obtain classification rules for attribution boundary labels from multi-dimensional classification rules. These classification rules include whether the knowledge content carried by the historical session context possesses knowledge reusability. Furthermore, deep learning models can be used to extract entity dependency features, scene binding features, and general knowledge features from the historical session context based on the attribution boundary label classification rules. Entity dependency features refer to whether the knowledge content contains personal entity information strongly bound to the target object, such as personal device serial numbers, personal resource identifiers such as server identifiers, container identifiers, and virtual machine identifiers. The presence of this type of information directly prevents the knowledge from being reused by other users. Scene binding features refer to whether the knowledge content contains personalized scene constraints applicable to the target object, such as "My cat's name is DouDou" or "My commonly used development environment configuration." These constraints limit the applicability of the knowledge to a specific individual and a specific scene, lacking cross-user universality. General knowledge characteristics refer to whether the knowledge content contains objective facts, general methods, industry standards, etc., that are independent of specific individuals or scenarios, such as "steps for troubleshooting memory faults," "Python list comprehension syntax," and "the principle of the Transmission Control Protocol (TCP) three-way handshake." This type of characteristic is the core indicator that knowledge has cross-user reusability.

[0052] Furthermore, deep learning models can be used to determine whether knowledge content possesses knowledge reusability based on entity dependency features, scene binding features, and general knowledge features. If the entity dependency feature indicates the existence of personal entity information bound to the target object, and / or the scene binding feature indicates the existence of personalized scene constraints applicable only to the target object, and / or the general knowledge feature indicates the absence of general knowledge, then the knowledge content is determined to lack knowledge reusability.

[0053] If the entity dependency feature indicates that there is no personal entity information bound to the target object, the scenario binding feature indicates that there is no personalized scenario constraint applicable only to the target object, and the general knowledge feature indicates that it contains general knowledge, then the knowledge content is determined to have knowledge reusability.

[0054] Furthermore, if the knowledge content lacks reusability, the ownership boundary label of the historical session context is determined to be a private boundary label. This private boundary label may include: the session identifier, user identifier, tenant identifier, and identifier of the digital human associated with the agent, etc. If the historical session context belongs to a group session, the private boundary label may also include: the group identifier to which the historical session context belongs.

[0055] Accordingly, if the knowledge content does not have knowledge reusability, the ownership boundary label of the historical session context can be determined as the global boundary.

[0056] This embodiment uses the reusability of the knowledge content itself as the criterion for determining the attribution boundary. Interactions involving personal entity information within group chats are automatically categorized as private boundaries, while interactions containing general knowledge within personal conversations are automatically categorized as global boundaries. This eliminates privacy risks and knowledge silo issues arising from misalignments between conversation format and content security attributes, ensuring the reliability of data isolation in enterprise-level scenarios. The abstract concept of knowledge reusability is broken down into semantic features such as entity dependency features, scenario binding features, and general knowledge features. This ensures that the attribution boundary determination process follows structured rules, allowing inputs with the same semantic features to produce the same boundary label. This avoids label drift and unreproducibility issues caused by freely generated models, providing a stable and reliable classification foundation for subsequent memory retrieval and access control.

[0057] In some embodiments, when the attribution boundary label is a private boundary, the multi-dimensional classification label further includes a task scenario label. Accordingly, when using a deep learning model to perform multi-dimensional classification of historical session contexts according to the multi-dimensional classification rules, the deep learning model can also obtain the classification rules for the task scenario label from the multi-dimensional classification rules. The classification rules for the scenario label include classifying application scenario types based on the temporal regularity of the target task associated with the knowledge content.

[0058] Accordingly, deep learning models can be used to extract task semantic features, instruction pattern features, and contextual clue features representing temporal regularity from historical conversation contexts based on the classification rules of task scenario labels. Among these, task semantic features refer to semantic vectors or structured labels parsed from historical conversation contexts that represent the essential attributes of the task performed in that interaction. They reflect what the dialogue was doing in the historical conversation context; for example, semantic matching can map "generating today's daily stock earnings report" to the steady-state task category of "daily earnings analysis," or "investigating the sudden network outage in server room 3" to the occasional task category of "emergency fault handling."

[0059] Steady-state tasks and occasional tasks are semantically distinct. Steady-state tasks typically correspond to fixed application or service processes and standardized outputs (such as daily reports, weekly reports, and periodic reviews), exhibiting high semantic stability and enumerability. In contrast, occasional tasks often correspond to non-standard and unconventional transactions, resulting in more random and specific semantic expressions. By pre-setting semantic libraries for steady-state and occasional tasks, or training models to classify semantics, task categories can be directly identified within a single session, thereby inferring their temporal regularity. Matching task semantic features to the steady-state task semantic library indicates high temporal regularity, while matching task semantic features to the occasional semantic library indicates low temporal regularity.

[0060] Command pattern features refer to the formal characteristics extracted from user commands in the context of historical sessions, representing whether the command's sentence structure, verb collocation, and parameter slots conform to a fixed template for periodic tasks. It reflects "how the user gives commands." For example, "Generate [Date][Application] Report" and "Update [Periodic][Metric] Data" are typical periodic command templates, while "Check this for me" and "Temporarily process that" are non-template-based free expressions.

[0061] When users perform routine, steady-state tasks, they often develop highly rigid instruction patterns due to efficiency and habit. These patterns are a direct projection of temporal regularity into language. Even if explicit time words like "daily" or "weekly" don't appear in a single conversation, a strong correlation with high temporal regularity is established as long as the instruction structure matches a periodic template. Conversely, instructions for sporadic tasks typically lack fixed templates, with loose sentence structures and ambiguous referents. This formal irregularity is an external manifestation of a lack of temporal regularity. This feature compensates for potential ambiguities in pure semantic classification, providing a formal criterion independent of content.

[0062] Contextual cue features refer to keywords, phrases, or pragmatic markers extracted from the context of historical conversations that can explicitly affirm or deny temporal regularity. They reflect whether there are signals in the dialogue that directly indicate regularity, including positive cues (such as "routine," "regular," "daily," or "as is") and negative cues (such as "temporary," "sudden," "urgent," "this time," "accidental," or "specialized").

[0063] Contextual clues are a direct basis for judgment. When a user actively uses occasional trigger words in a single session, it's equivalent to explicitly declaring that the task lacks temporal regularity. In this case, regardless of the task semantics or instruction pattern, it can be preferentially classified as an occasional task scenario, avoiding misjudgment. Similarly, positive clues can directly confirm daily steady-state task scenarios. This feature serves as a verification and fallback mechanism for the first two features, ensuring the completeness of the classification logic.

[0064] Furthermore, based on task semantic features, instruction pattern features, and contextual clue features, the temporal regularity of the target task is determined. If the task semantic features match a pre-defined steady-state task semantic library, the instruction pattern features indicate that the task conforms to a periodic execution instruction template, and the contextual clue features indicate that the task does not contain occasional trigger words, then the target task associated with the knowledge content is determined to have temporal regularity. Furthermore, the task scenario tag can be determined to be a steady-state task scenario tag, which can be marked for permanent retention.

[0065] If the task semantic features match the preset sporadic task semantic library, and / or the contextual clue features indicate the presence of sporadic trigger words, then it is determined that the target task associated with the knowledge content does not have temporal regularity. Furthermore, the task scenario label can be determined as a sporadic task scenario label, which can be marked as eligible for overall archiving after the target task is completed. This scenario-level storage isolation prevents sporadic task data from polluting the long-term accumulated data of steady-state tasks.

[0066] If the target task exhibits temporal regularity, the task scenario label is determined as a steady-state task scenario label and can be permanently retained. If the target task does not exhibit temporal regularity, the task scenario label is determined as an occasional task scenario label.

[0067] This embodiment combines task semantic features, instruction pattern features, and contextual clue features. Instruction pattern features fill the formal blind spots in task semantic feature classification, while contextual clue features serve as a fallback check by directly accepting explicit user declarations. The combination of these three features ensures that even in the absence of explicit time words, semantic ambiguity, or non-standardized instructions, the temporal regularity can still be accurately inferred, thereby improving the robustness and accuracy of task scenario classification.

[0068] In some embodiments, when the attribution boundary label is a private boundary, the multi-dimensional classification label further includes: content attribute labels. Accordingly, when using a deep learning model to perform multi-dimensional classification of the historical conversation context according to multi-dimensional classification rules, the deep learning model can also obtain the classification rules for content attribute labels from the multi-dimensional classification rules. The classification rules for content attribute labels include dividing content attribute labels based on the specific semantics carried by the knowledge content. Further, the deep learning model can extract semantic features of the knowledge content according to the classification rules for content attribute labels. Specifically, the deep learning model can extract semantic features from the historical conversation context to extract semantic features representing the essential attributes of the knowledge content. These semantic features integrate multi-source signals such as text surface semantics, dialogue role information, and contextual information.

[0069] The deep learning model uses classification rules based on the semantic features and content attribute tags to determine the content attribute tags of the historical session context. Specifically, the deep learning model can sequentially or in parallel evaluate the confidence of semantic features belonging to multiple content attribute tags according to the matching logic defined in the content attribute tag classification rules. Further, the content attribute tags of the historical session context can be determined based on the confidence of semantic features belonging to multiple content attribute tags. For example, the content attribute tag with the highest confidence can be determined as the content attribute tag of the historical session context from multiple content attribute tags; or, the content attribute tag with a confidence greater than or equal to a set confidence threshold can be determined as the content attribute tag of the historical session context from multiple content attribute tags.

[0070] In this embodiment, based on the defined data visibility range by the ownership boundary, the content attribute tags further layer the private memory according to its semantic essence. Subsequently, the memory at the corresponding level can be retrieved in a targeted manner according to the content attribute type of the user interaction request, instead of traversing all private data, which can improve recall efficiency and relevance.

[0071] Specifically, if semantic features indicate that the knowledge content reflects personalized behavioral guidelines, then the content attribute label is determined to be a preference profile label. If semantic features indicate that the knowledge content reflects objective domain cognition, then the content attribute label is determined to be a factual knowledge label. If semantic features indicate that the knowledge content reflects a task execution trajectory that conforms to the application semantics of an intelligent agent, then the content attribute label is determined to be a task execution label. If semantic features indicate that the knowledge content reflects a decision-making deduction chain, then the content attribute label is determined to be a thinking logic label; if the semantic features indicate that the knowledge content reflects non-task-oriented corpus, then the content attribute label is determined to be a non-task interaction label.

[0072] This embodiment establishes a fine-grained semantic discrimination standard by defining clear semantic feature pointers for five types of content attributes, which helps to eliminate fuzzy boundaries between content attributes and provides an unambiguous indexing basis for subsequent differentiated processing.

[0073] In some embodiments, when the attribution boundary label is a private boundary, the multi-dimensional classification label further includes a storage type label. Accordingly, when using a deep learning model to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules, the deep learning model can also obtain the classification rules for the storage type label from the multi-dimensional classification rules; the classification rules for the storage type label include whether the historical session context contains resource-intensive files. Further, the deep learning model can be used to traverse the data types of the data contained in the historical session context; if resource-intensive files exist among the data types contained in the historical session context, then the storage type label of the historical session context is determined to be a compressed storage label. If resource-intensive files do not exist among the data types contained in the historical session context, then the storage type label of the historical session context is determined to be a plain text storage label.

[0074] This solution automatically identifies the data types in the context of historical sessions using a deep learning model, and dynamically generates storage type labels based on whether resource-intensive files are included. Its technical effectiveness is mainly reflected in the following three aspects: In this embodiment, the accurate identification of data form is completed during the classification stage. Subsequently, the session context containing resource-intensive files is automatically marked as compressed storage tag, and the session context of plain text or lightweight data is marked as original text storage tag. The original structure is preserved to support efficient retrieval and incremental updates, realizing adaptive allocation of storage resources and reducing the storage cost of long-term memory.

[0075] The foregoing embodiments exemplify how to perform multi-dimensional classification of historical session context using an agent-dependent deep learning model based on multi-dimensional classification rules, but this is not intended to be limiting. In other embodiments, a hybrid pipeline approach combining a rule engine and a lightweight dedicated model can be used to perform multi-dimensional classification of historical session context to determine a first multi-dimensional classification label. Specifically, the rule engine directly matches the structured metadata in the historical session context with the explicit features in the multi-dimensional classification rules. For example, for the attribution boundary dimension, explicit values ​​of entity dependency features and scene binding features can be directly extracted from fields such as session identifier, user identifier, and group identifier in the historical session context; for the storage type dimension, the data objects contained in the context can be traversed through a file type detection interface to determine whether resource-intensive files exist. This type of determination does not require semantic reasoning, is based on predefined Boolean logic or regular expression matching, and the output result is deterministic, avoiding the illusion that large models may produce in simple field recognition.

[0076] For dimensions that cannot be directly determined by explicit rules, such as preference profiles and thinking logic in the content attribute dimension, general knowledge features requiring semantic understanding in the attribution boundary, and task semantic features and instruction pattern features in the application scenario dimension, a pre-trained lightweight dedicated classification model is invoked for processing. One dedicated classification model corresponds to one classification dimension. The dedicated classification model can be a multi-label text classifier fine-tuned based on Bidirectional Encoder Representations from Transformers (BERT), an intent recognition model based on fasttext, or a retrieval enhancement model based on keyword-semantic joint matching, etc. The input to the dedicated classification model includes: relevant text fragments and corresponding classification hints for the current dimension to be classified, and the output is the candidate label and confidence score for that dimension. Since each dedicated classification model is responsible for a single-dimensional classification task, its parameter count and inference overhead are far lower than general-purpose large models, and it can be finely tuned for specific domain corpora, resulting in more stable classification accuracy.

[0077] Furthermore, the result aggregation module can combine and verify the explicit labels output by the rule engine with the implicit labels output by the dedicated classification model. If there is a conflict between the two results of a certain dimension (e.g., the rule engine determines it as a global boundary but the semantic model determines it as a private boundary), arbitration is performed according to a preset priority strategy (usually the rule engine takes precedence) or a confidence threshold to determine the final classification label for that dimension. The labels corresponding to multiple classification dimensions are combined to obtain multi-dimensional classification labels for the historical session context. The entire pipeline supports parallel execution of the judgment tasks for each dimension, which can reduce the overall latency. At the same time, each dedicated classification model is deployed and iterated independently. When the classification rules for a certain dimension change, only the corresponding dedicated classification model needs to be modified, without retraining or replacing the entire deep learning model, which has stronger operability and scalability.

[0078] After determining the multi-dimensional classification labels of the historical conversation context in step 102, the historical conversation context and the multi-dimensional classification labels can be associated and stored in step 103. Specifically, the multi-dimensional classification labels of the historical conversation context can be used as the metadata index key, and the historical conversation context can be used as the data payload. The two are persistently stored in the form of key-value pairs or structured records. This storage structure supports targeted retrieval by any combination of dimensional labels. For example, the target memory can be quickly located by "the ownership boundary is a private boundary and the content attribute is a preference profile" without having to traverse the entire original dialogue text.

[0079] Simultaneously, the stored procedure can automatically adapt the write strategy based on the storage type field in the multi-dimensional label. Specifically, if the storage type label in the multi-dimensional storage label is a compressed storage label, then resource-intensive files in the historical session context can be separated to object storage, and only the reference path of the resource-intensive file in object storage can be retained. Other data in the historical session context besides resource-intensive files, along with the multi-dimensional category label, can be stored in the main index. Alternatively, if the storage type label in the multi-dimensional storage label is a compressed storage label, then the historical session context can be compressed according to the set compression algorithm, and the compressed historical session context can be associated and stored with the multi-dimensional category label.

[0080] If the storage type label in the multi-dimensional storage tag is the original text storage label, then the historical session context is directly associated with the multi-dimensional classification label for storage. This associated storage mechanism transforms the unstructured session context into an indexable and manageable structured memory asset, providing underlying data support for subsequent accurate retrieval based on multi-dimensional classification labels.

[0081] This application embodiment utilizes a multi-dimensional classification tagging mechanism to store the multi-dimensional classification tags and associations of the historical session context between the agent and the target object, constructing a multi-level memory organization architecture. This enables the agent to quickly recall information related to the user's interaction request from the historical session context based on the multi-dimensional classification tags in subsequent interactions. This breaks through the lifecycle of a single session and allows for the effective reuse of knowledge from historical session content through multi-dimensional classification tag matching across different sessions. This cross-session knowledge reuse and context continuation helps enhance the agent's long-term memory capabilities and promotes the evolution of AI systems such as agents from tools for performing isolated tasks to intelligent assets that carry organizational experience and wisdom.

[0082] Based on the associated storage of historical session context and multi-dimensional classification labels, for subsequent user interaction requests received by the agent, the target information required for the user interaction request can be retrieved from the stored historical session context based on the multi-dimensional classification labels. The following section combines... Figure 2 This paper details the specific implementation method for retrieving target information required for user interaction requests from stored historical session contexts based on multi-dimensional classification labels. Specifically, the step of retrieving the target information based on multi-dimensional classification labels from stored historical session contexts is performed asynchronously with the aforementioned multi-dimensional classification and storage of historical session contexts. Correspondingly, the device, apparatus, or module (such as the agent) executing the agent-oriented information management method can also be executed asynchronously. Figure 2Based on multi-dimensional classification labels, the system retrieves relevant operations for the target information required by user interaction requests from the stored historical session context. Specifically, the historical session context of at least one of the agent's historical sessions is associated with and stored along with its corresponding multi-dimensional classification labels.

[0083] Figure 2 This is a flowchart illustrating another agent-oriented information management method provided in an embodiment of this application. This information management method is primarily applied to information retrieval. For example... Figure 2 As shown, this information management method mainly includes the following steps: 201. Obtain user interaction requests.

[0084] 202. Perform semantic analysis on the user interaction request to obtain the interaction intent reflected by the user interaction request.

[0085] 203. Determine the user boundary identifier for the current interaction based on the session identifier corresponding to the user interaction request.

[0086] 204. Map user boundary identifiers and interaction intents to multi-dimensional category labels to determine the second multi-dimensional category label corresponding to the user interaction request.

[0087] 205. From the historical session context of at least one historical session, filter out the target historical session context associated with the second multi-dimensional classification label.

[0088] 206. Based on the interaction intent, retrieve information from the target historical session context to obtain the target information required by the user's interaction request.

[0089] In this embodiment, before acquiring the user interaction request, the agent's historical session context from at least one previous session is associated and stored with the multi-dimensional classification labels of that historical session context. The user interaction request in step 201 is the user interaction request currently received by the agent. The user interaction request can be described using natural language, such as text, voice, or multimodal commands, or it can be described using a scripting language.

[0090] User interaction requests are bound to session metadata, including session identifier, user identifier, group identifier (if present), digital human identifier, and request timestamp. User interaction requests not only convey the content explicitly expressed by the user, but also implicitly carry key contextual clues for subsequent boundary determination and context association, which can avoid boundary misjudgment or intent ambiguity due to missing metadata.

[0091] After obtaining the user interaction request, the semantic understanding stage can be entered, transforming the raw input into a computable structured intent representation, providing refined input for subsequent label mapping. Accordingly, in step 202, the user interaction request can be semantically parsed to obtain the interaction intent reflected by the user interaction request.

[0092] Specifically, deep learning models or dedicated intent recognition modules that the agent relies on can be invoked to parse core actions, entity objects, content attribute requirements, and implicit task scenario expectations from user interaction requests. For example, "Help me check my commonly used development environment configuration" can be parsed as "Retrieve personal preferences + development tools involved + private boundaries". This interaction intent is no longer a one-dimensional classification label, but a multi-dimensional semantic vector containing content attribute preferences, task scenario hints, and boundary sensitivity.

[0093] After completing the semantic parsing of intent, in step 203, the user boundary identifier for the current interaction can be determined based on the session identifier corresponding to the user interaction request. Specifically, the user identifier, group identifier, tenant identifier, and digital human identifier can be extracted from the session metadata corresponding to the user interaction request, and combined to generate the user boundary identifier according to the type (individual / group) and affiliation of the current session. For example, in an individual session, the user boundary identifier may include the user identifier and the digital human identifier; in a group session, the user boundary identifier may include the group identifier, user identifier, digital human identifier, and tenant tag. This user boundary identifier for the current interaction is not fixed but dynamically generated according to the session context of each user interaction request, reflecting the visible range of the current interaction subject.

[0094] Interaction intent and user boundary identifiers constitute the request semantics of a user interaction request. Subsequently, these can be mapped to a multi-dimensional classification label system aligned with the historical memory storage structure. Accordingly, in step 204, user boundary identifiers and interaction intents can be mapped to multi-dimensional classification labels to determine the multi-dimensional classification label corresponding to the user interaction request. For ease of description and differentiation, [the following is a simplified explanation of the process]. Figure 1 The related embodiments define the multi-dimensional classification labels obtained by multi-dimensionally classifying the historical session context as the first multi-dimensional classification labels; and define the multi-dimensional classification labels corresponding to the user interaction requests in step 204 as the second multi-dimensional classification labels. Step 204 is the core bridge connecting user interaction requests and historical memory, transforming dynamic request semantics into structured retrieval labels (i.e., the second multi-dimensional classification labels) that are fully aligned with the storage structure.

[0095] In some embodiments, user boundary identifiers can be converted into attribution boundary labels in multi-dimensional classification labels to obtain target attribution boundary labels. Specifically, the identifier fields in the attribution boundary label can be assembled according to the organizational structure of attribution boundary labels in multi-dimensional classification labels to obtain the target attribution boundary label. This conversion does not require semantic reasoning and ensures that the organizational structure of the attribution boundary label on the request side is consistent with that on the storage side through deterministic field mapping.

[0096] Furthermore, scenario intent features representing scenario needs and content intent features representing content needs can be extracted from interaction intents. Specifically, through predefined semantic slots, lightweight feature extractors, or agent-dependent deep learning models, task scenario intent features representing the temporal regularity of tasks (such as semantic signals like "daily," "routine," "temporary," and "sudden") and content intent features representing content needs can be extracted from interaction intents. Scenario intent features refer to semantic signals extracted from user interaction intents that represent whether the task associated with the user's interaction request has regularity, periodicity, or stability in the time dimension. It reflects the type of task scenario to which the user expects to access the memory, rather than the specific content of the task itself.

[0097] Scene intent features serve as a crucial bridge connecting dynamic user requests with the hierarchical structure of steady-state and occasional tasks in historical memory. They determine whether a current user interaction request should point to long-accumulated daily experience or a temporary, specific task. For example, when a user asks, "What is the logic behind the daily stock market analysis?", the terms "daily" and "analysis logic" are typical scene intent features, exhibiting temporal regularity. Conversely, when a user says, "Help me recall the process of handling the network outage in the server room last Wednesday," "last Wednesday" and "network outage handling" constitute scene intent features lacking temporal regularity. Scene intent features do not focus on what is analyzed or handled, but only on whether the task repeats at a fixed rhythm, thus providing a basis for the agent to select a subset of steady-state or occasional task memories.

[0098] Content intent features are signals extracted from interaction intent that characterize the semantically essential content attribute of the knowledge required for a request. They reflect the nature of the knowledge the user wants to acquire, rather than the knowledge's boundaries or task context. Content intent features are the core basis for accurately locating specific semantic levels (i.e., content attribute levels) within historical memory, directly corresponding to the user's content needs for knowledge types. Their core value lies in transforming vague natural language requests into indexable content attribute tags. For example, if a user asks, "What editor do I like to use?", the words "like" and "use" constitute a preference profile-type content intent feature; if a user asks, "What is the principle of the TCP three-way handshake?", the words "principle" and "TCP" constitute a factual knowledge-type content intent feature; and if a user asks, "Why did you decide to restart the service instead of expanding its capacity last time?", the words "why" and "decided" constitute a logical reasoning-type content intent feature. Content intent features focus on the user's essential need for knowledge content, avoiding interference from task names or boundary information, and ensuring that the retrieved content aligns with the user's actual information needs.

[0099] Furthermore, the scene intent features can be converted into task scene labels within a multi-dimensional classification system to obtain the target task scene label. Specifically, the extracted scene intent features can be matched against pre-defined steady-state task semantic libraries and occasional task semantic libraries. If the scene intent features match the steady-state semantic library, the target task scene label is determined to be a steady-state task scene label; otherwise, it is determined to be an occasional task scene label. Alternatively, the scene intent features can be input into a deep learning model relied upon by the agent. The deep learning model then determines the task scene type based on the scene intent features and outputs the corresponding target task scene label.

[0100] Correspondingly, the content intent features can be converted into content attribute tags in multi-dimensional classification tags to obtain the target content attribute tags. In some embodiments, the semantic similarity between the content intent features and the semantic definitions of multiple content attribute tags can be calculated to obtain the semantic similarity between the content intent features and the semantic definitions of multiple content attribute tags. From the multiple content attribute tags, the content attribute tags with a semantic similarity greater than or equal to a set similarity threshold are determined as the target content attribute tags.

[0101] Furthermore, the target boundary label, the target task scenario label, and the target content attribute label can be combined into a second multi-dimensional classification label.

[0102] By directly converting user boundary identifiers into attribution boundary labels, potential boundary misjudgments during model inference are avoided. Subsequent filtering can then use these attribution boundary labels for forced filtering, ensuring that only memories accessible to the current user are recalled, reducing the risk of data leakage across users and tenants. By extracting scene-specific intent and content-specific intent from interaction intents, the intelligent agent can simultaneously match both task scene and content attribute dimensions. This dual-dimensional localization further narrows the search scope, reduces invalid traversals, and improves the relevance and response speed of the retrieved results.

[0103] In other embodiments, an end-to-end semantic mapping model is used as an alternative implementation. Specifically, the joint representation learning capability of a deep learning model can be utilized to directly map user boundary identifiers and interaction intentions into structured second multi-dimensional classification labels. Specifically, user boundary identifiers and interaction intentions (such as "Why was pandas recommended last time?") can be input into a deep learning model dependent on the agent; the deep learning model can assemble the identifier fields in the attribution boundary label according to the organizational structure of the attribution boundary labels in the multi-dimensional classification labels to obtain the target attribution boundary label. The deep learning model can also determine the similarity between the interaction intention and multiple content attribute labels, as well as the similarity between the interaction intention and multiple task scenario labels, based on the interaction intention. Then, based on the similarity between the interaction intention and multiple content attribute labels, content attribute labels with a similarity greater than or equal to a set similarity threshold can be determined as target content attribute labels. Correspondingly, task scenario labels with higher similarity can also be determined as target task scenario labels from among the task scenario labels. Furthermore, the deep learning model can combine the target boundary label, the target task scene label, and the target content attribute label into a second multi-dimensional classification label, and output the second multi-dimensional classification label.

[0104] This embodiment utilizes an agent-dependent deep learning model to capture implicit semantic dependencies in interactive intents. The deep learning model can simultaneously infer private boundaries, steady-state task scenarios, and preference profile content attributes, avoiding label loss due to incomplete feature decoupling during step-by-step transformation. The deep learning model is more robust to ambiguous, omitted, or non-standardized expressions. Even if the interactive intent does not explicitly mention temporal regularity or content type, the deep learning model can still jointly infer based on context and boundary information, improving the robustness of determining multi-dimensional classification labels.

[0105] After determining the second multi-dimensional classification label, in step 205, the target historical session context associated with the second multi-dimensional classification label can be filtered from the historical session context of at least one pre-stored historical session.

[0106] Step 205 can use the second multi-dimensional classification tags as index keys to perform efficient targeted retrieval within the associated stored historical session context. The second multi-dimensional classification tags include: target affiliation boundary tags, target task scenario tags, and target content attribute tags.

[0107] Accordingly, a first historical session context with the target attribution boundary label can be selected from the historical session contexts of at least one historical session. Furthermore, a second historical session context with the target content attribute label can be selected from the first historical session context as a candidate historical session context.

[0108] In some embodiments, a second historical session context whose content attribute tag matches the target content attribute tag can be selected from the first historical session context as a candidate historical session context. This embodiment filters historical session contexts corresponding to the target content attribute tag from the historical session contexts where the current user has permissions. When the user's intent clearly points to a specific knowledge type, only information matching the content attribute is returned, which can reduce interference caused by redundant information and reduce the context length and token consumption in subsequent reasoning stages.

[0109] In other embodiments, when the target content attribute tag is not a preference profile tag, a second historical session context with the target content attribute tag is selected from the first historical session context; and a third historical session context with the preference profile tag is selected from the first historical session context; the second and third historical session contexts are then determined as candidate historical session contexts. This embodiment automatically supplements preference profile tags when the target content attribute is not a preference profile tag. While returning the target content memory, it appends preference profile-type context, providing the agent with background information on personalized user behavior guidelines. This allows the agent to automatically adapt to the user's expression habits, tool preferences, or interaction style when organizing responses, improving the coherence of the response and the user experience.

[0110] After selecting candidate historical session contexts, the target historical session context can be selected from the candidate historical session contexts based on the target task scenario tag. This embodiment adopts a three-level filtering mechanism based on attribution boundary, content attribute (i.e., task scenario), which can reduce the traversal overhead of invalid data by gradually converging the retrieval scope. Using the target attribution boundary tag as a mandatory filtering condition, all memories outside the current user's permission range are excluded, avoiding subsequent computational resources being wasted on unauthorized data. In the historical session contexts where the current user has permissions, the semantically matching subset of memories is quickly located by content attribute tags, skipping content levels irrelevant to the current requested knowledge type; the final filtering based on task scenario tags reduces the traversal scope and improves the overall retrieval response speed.

[0111] Specifically, if the target task scenario label is a steady-state task scenario, then from the candidate historical session contexts, historical session contexts with the same task scenario label as a steady-state task scenario are selected as the target historical session context. When the target task scenario is a steady-state task scenario, retaining historical session contexts with the same task scenario label as a steady-state task scenario can eliminate interference from occasional task memories, prevent temporary and fragmented data from polluting the output quality of steady-state services, and ensure the reliability and predictability of the agent in daily tasks.

[0112] If the target task scenario is labeled as an occasional task scenario, then from the candidate historical session contexts, the fourth historical session context labeled as a steady-state task scenario and the fifth historical session context labeled as an occasional task scenario are selected, and the fourth and fifth historical session contexts are determined as the target historical session contexts. In this embodiment, when the target task scenario is an occasional task, both the historical session contexts of the occasional task scenario and the historical session contexts of the steady-state task scenario are recalled simultaneously, achieving cross-task scenario information retrieval. The occasional task memory provides specific details of this particular event, while the steady-state task memory contributes general processing flows, domain knowledge, or similar historical cases. The two complement each other to form a complete problem-solving basis, preventing knowledge gaps caused by only recalling occasional memories.

[0113] After determining the target historical session context, in step 206, information retrieval can be performed within the target historical session context based on the interaction intent to obtain the target information required for the user's interaction request. This embodiment transforms unstructured dialogues generated by the user in historical sessions into structured memory assets with multi-dimensional classification labels through intent parsing and multi-dimensional classification label mapping. When a new session initiates a user interaction request, the agent can index relevant knowledge fragments from historical sessions based on standardized multi-dimensional labels, enabling knowledge to transcend time spans and physical isolation of sessions, achieving cross-session knowledge reuse and helping to enhance the agent's long-term memory capabilities.

[0114] On the other hand, this embodiment uses user boundary identifiers as input for multi-dimensional classification label mapping, ensuring that the attribution boundary in the second multi-dimensional classification label is consistent with the permissions of the current interaction subject. Subsequent filtering can directly use this attribution boundary label for pre-filtering, thereby preventing the risk of data privilege escalation across users and tenants, while avoiding the performance overhead of full traversal. Using multi-dimensional classification label retrieval can quickly skip irrelevant historical session contexts, reduce the size of the recalled dataset, and accelerate knowledge retrieval.

[0115] The following is an exemplary description of a specific implementation method for retrieving target information from the target historical session context based on the interaction intent to obtain the target information required for the user's interaction request. In some embodiments, such as Figure 3 As shown in step S1, semantic feature extraction and keyword feature extraction can be performed on the interaction intent to obtain the interaction semantic vector and the first keyword feature. Correspondingly, as... Figure 3 As shown in step S2, the target historical session context can also be segmented to obtain multiple candidate session fragments; and as shown in step S2... Figure 3 As shown in step S3, semantic feature extraction and keyword feature extraction are performed on multiple candidate conversation segments respectively, resulting in segment semantic vectors and second keyword features corresponding to each candidate conversation segment. The semantic feature extraction model for semantic feature extraction of interaction intent is the same as the model for semantic feature extraction of candidate conversation segments. The semantic feature extraction model can be a dense embedding model, such as the BERT model. Similarly, the keyword extraction model for keyword feature extraction of interaction intent is the same as the model for keyword feature extraction of candidate conversation segments. For example, the same Best Matching 25 (BM25) model is used. BM25 is a statistical keyword retrieval algorithm that scores the relevance between query keywords and documents based on statistical features such as term frequency and inverse document frequency.

[0116] Furthermore, such as Figure 3 As shown in step S4, the first similarity between the interaction intent and each candidate session segment can be determined based on the interaction semantic vector and the segment semantic vectors corresponding to each of the multiple candidate session segments. For example, for any candidate session segment X, the cosine similarity between the interaction semantic vector and the segment semantic vector of candidate session segment X can be calculated as the first similarity.

[0117] Accordingly, such as Figure 3As shown in step S5, a second similarity between the interaction intent and each candidate conversation segment can be determined based on the first keyword feature and the second keyword features corresponding to each of the multiple candidate conversation segments. Specifically, the BM25 algorithm can be used to calculate the similarity score between the first keyword feature and the second keyword feature corresponding to the candidate conversation segment X, which is then used as the second similarity.

[0118] Furthermore, the target session segment can be determined from multiple candidate session segments based on the first similarity and the second similarity.

[0119] In some embodiments, such as Figure 3 As shown in step S6, for any candidate session fragment X, the first similarity and second similarity corresponding to candidate session fragment X can be weighted and fused to obtain the baseline recall score of candidate session fragment X. For example, according to the weights corresponding to the first similarity and second similarity, the first similarity and second similarity corresponding to candidate session fragment X can be weighted and summed to obtain the baseline recall score of candidate session fragment X. Using the same method, the similarity between each candidate session fragment and the interaction intent can be obtained. Then, from multiple candidate session fragments, candidate session fragments with a baseline recall score greater than or equal to a set score threshold can be selected as target session fragments.

[0120] In other embodiments, such as Figure 3 As shown in step S7, the time attributes and historical recall attributes corresponding to each of the multiple candidate session fragments can also be obtained. The time attribute of a candidate session fragment is a quantitative indicator characterizing its timeliness and freshness, and may include the time difference between the generation time of the candidate session fragment and the current time. The time attribute reflects the time span of the remembered content from the present, and is the core basis for judging whether knowledge is outdated, whether it still has reference value, and whether decay needs to be triggered. The smaller the time difference, the fresher the candidate session fragment is, and usually has a higher immediate relevance; the larger the time difference, the more outdated the content of the candidate session fragment may be, and its reuse value may be lower.

[0121] Historical recall attributes are quantitative indicators that characterize the frequency of use and recent activity of candidate session fragments in past interactions. They may include: the time when the candidate session fragment was last recalled and the cumulative number of times the candidate session fragment has been recalled. The time when the candidate session fragment was last recalled reflects its short-term activity status and is used to identify hot knowledge that is continuously in demand; the cumulative number of recalls reflects the long-term reuse value of the candidate session fragment, and high-frequency recall usually verifies the stability and irreplaceability of the candidate session fragment in a specific scenario.

[0122] Furthermore, such as Figure 3As shown in step S8, for any candidate session segment X among multiple candidate session segments, the baseline recall score of candidate session segment X is weighted and corrected according to the time attribute and historical recall attribute of candidate session segment X to obtain the target recall score of candidate session segment X.

[0123] Specifically, the time aging factor of candidate session fragment X can be calculated based on its time attributes. For example, the time difference between the generation time of candidate session fragment X and the current time (e.g., in days) can be substituted into a preset time aging decay function. .in, As a time aging factor, Let X be the time difference between the generation time of candidate session fragment X and the current time. half_life For example, the pre-configured half-life parameter. half_life It can be set to 90 days, 120 days, etc. The time aging factor causes the target recall score to decrease exponentially over time, ensuring that the weight of old memories automatically decays and preventing outdated knowledge from interfering with the current response.

[0124] Accordingly, a recent hit weighting factor for candidate session fragment X can be calculated based on the most recent recall time of the candidate session fragment X. Specifically, it can be determined whether the time difference between the most recent recall time of candidate session fragment X and the current time is less than or equal to a set time difference threshold, such as 7 days, 10 days, etc.: If the time difference between the most recent recall time of candidate session fragment X and the current time is less than or equal to the set time difference threshold, then the recent hit weighting factor for candidate session fragment X is a preset first recovery coefficient. The first recovery coefficient is greater than 1, such as 1.5 or 2, etc., used to compensate for the recall score loss caused by time aging, ensuring that recently used memories maintain a high priority; if the time difference between the most recent recall time of candidate session fragment X and the current time is greater than the set time difference threshold, then the recent hit weighting factor for candidate session fragment X is a preset second recovery coefficient. The second recovery coefficient is less than the first recovery coefficient, such as the second recovery coefficient being equal to 1, and no additional weighting is applied. The recent hit weighting factor for candidate session fragment X can identify short-term active memories, preventing the erroneous rejection of continuously valuable knowledge due to time decay.

[0125] Accordingly, a reference frequency weighting factor for candidate session fragment X can be calculated based on the cumulative recall count of candidate session fragment X. Specifically, the cumulative recall count of candidate session fragment X is substituted into a monotonically increasing objective function. The objective function can be: .in, frequency bonus As a citation frequency weighting factor, hit_count To accumulate the number of recalls, kThis is a normalization constant used to control the magnitude of the bonus. The citation frequency weighting factor can give frequently reused candidate session fragments a long-term retention bonus, verifying the stability and irreplaceability of the candidate session fragment in historical interactions, and avoiding the premature obsolescence of high-value accumulated knowledge due to reliance solely on timeliness.

[0126] Furthermore, the baseline recall score of candidate session fragment X can be multiplied by its time aging factor, recent hit weighting factor, and citation frequency weighting factor to obtain the target recall score of candidate session fragment X. The target recall score of candidate session fragment X can be expressed as: . base_score Represents the baseline recall score for candidate session fragment X; recency_boost This represents the most recent hit weighting factor for candidate session fragment X. This target recall score integrates four dimensions: content relevance, timeliness, short-term activity, and long-term reuse value. As a basis for subsequent knowledge recall, it can ensure that the recall results not only match the current interaction intent but also take into account the life cycle status of memory and actual utility.

[0127] Specifically, the time aging factor automatically reduces the weight of outdated content to prevent obsolete knowledge from interfering with the response; the recent hit weighting restores the priority of short-term active memories to prevent continuously useful hot knowledge from being mistakenly rejected; and the citation frequency bonus strengthens the long-term value of frequently reused memories, ensuring that validated and stable knowledge is not prematurely eliminated by the timeliness mechanism. These three factors work together to distinguish between memories in different states, such as fresh but useless, outdated but high-frequency, and recently active and relevant, enabling dynamic optimization of recall results.

[0128] Furthermore, such as Figure 3 As shown in step S9, a target session segment (defined as the first target session segment) can be determined from multiple candidate session segments based on their respective target recall scores. For example, candidate session segments with target recall scores greater than or equal to a set score threshold can be determined from multiple candidate session segments and used as the first target session segment.

[0129] This embodiment uses a weighted fusion of the first and second similarity scores as the baseline recall score. This avoids the limitations of a single similarity metric in complex queries and provides a reliable initial anchor for subsequent recall score adjustments, preventing highly relevant memories from being prematurely eliminated due to initial recall score bias. Furthermore, by weighting the baseline score with time and historical recall attributes, the target recall score no longer relies solely on static content matching but reflects the lifecycle status and actual utility of the memory in real time. This enables dynamic adaptive evaluation of memory value, contributing to improved accuracy of the recall results.

[0130] After selecting the first target conversation segment, target information can be determined based on it. This embodiment simultaneously extracts semantic vectors and keyword features from both the interaction intent and candidate conversation segments. This allows the recall process to capture deep semantic relationships such as synonyms and implicit intentions through the first similarity score, and to anchor keywords through the second similarity score. This dual-channel complementarity avoids the problems of pure semantic retrieval mismatch with key entities and pure keyword retrieval failing to generalize semantics, thus improving the comprehensiveness and accuracy of candidate conversation segments.

[0131] On the other hand, segmenting the target historical session context into blocks and breaking down long texts into multiple candidate session fragments allows similarity calculations to focus on local semantic units rather than the entire context, which can reduce noise interference in single feature extraction and matching.

[0132] In some embodiments, before determining the target information based on the first target session fragment, such as Figure 3 As shown in step S10, a relevance score between the first target session segment and the user interaction request can also be calculated. If the relevance score is greater than or equal to a set score threshold (i.e., ...), the relevance score is calculated. Figure 3 If step S11 determines the result as yes, then proceed directly. Figure 3 Step S12 shows the operation of determining target information based on the first target session fragment. In this embodiment, Figure 3 The “currently recalled target session fragment” in step S12 is the first target session fragment.

[0133] In some embodiments, if the first target session fragment comes from the historical session context of the same content attribute tag, then the first target session fragment can be identified as the target information.

[0134] In other embodiments, there are multiple first target session fragments. These multiple first target session fragments originate from historical session contexts corresponding to multiple content attribute tags. Accordingly, the multiple first target session fragments can be sorted according to the recall priority of the multiple content attribute tags to obtain target information. The recall priority among content attribute tags can be implemented as follows: preference profile tags > factual knowledge tags > task execution tags > thinking logic tags > non-task interaction tags, ensuring that high-value memories are prioritized.

[0135] For example, a user says "I like pandas" in the first session, "Help me process CSV" in the fifth session, and asks "Why do you recommend pandas?" in the tenth session. For the tenth session, the target session fragments recalled from the first two sessions can be sorted according to recall priority to obtain the target information: "Because you said you liked pandas before, and you have processed large amounts of CSV data, I recommended pandas based on your preferences and experience."

[0136] This embodiment introduces the recall priority of content attribute tags to sort and integrate multiple target session fragments from different historical session contexts. It recalls relevant fragments from the historical session contexts corresponding to different content attribute tags to achieve cross-session knowledge integration, so that the final output target information is no longer a single-dimensional fragment, but a cross-session and cross-content attribute knowledge collaborative reuse.

[0137] In other embodiments, if the relevance score between the aforementioned first target session fragment and the user interaction request is less than a set score threshold (i.e. Figure 3 If step S11 results in a negative result, it indicates that the currently recalled information is insufficient to respond to the user's interaction request. In this case, other target session fragments associated with the interaction intent (defined as second target session fragments) can be recalled in the first historical session context corresponding to other content attribute tags. Other content attribute tags refer to content attribute tags other than those in the target historical session context.

[0138] Accordingly, if the relevance score between the first target session fragment and the user interaction request is less than a set score threshold (i.e., Figure 3 If step S11 determines the result as negative, then... Figure 3 As shown in step S13, the content attribute tag to be recalled can be determined from the currently unrecalled content attribute tags based on the recall priority of multiple content attribute tags. For example, from multiple content attribute tags, the currently unrecalled content attribute tag is determined; from the currently unrecalled content attribute tags, the content attribute tag with the highest recall priority is determined as the content attribute tag to be recalled. Then, as... Figure 3 As shown in step S14, in the first historical session context corresponding to the content attribute tag to be recalled, the second target session fragment associated with the interaction intent is recalled. For the specific implementation of recalling the second target session fragment in the first historical session context corresponding to the content attribute tag to be recalled, please refer to the aforementioned content on recalling the first target session fragment in the target historical session context, which will not be repeated here.

[0139] Furthermore, such as Figure 3 As shown in step S15, the first target session fragment can be combined with the currently recalled second target session fragment to form the currently recalled target session fragment. Then, as... Figure 3 As shown in step S16, a relevance score between the currently recalled target session segment and the user interaction request can also be calculated. If the relevance score between the currently recalled target session segment and the user interaction request is greater than or equal to a set score threshold (i.e., ... Figure 3 If the result of step S11 is yes, then... Figure 3As shown in step S12, the target information required for the user interaction request is determined based on the currently recalled target session fragment. For the specific implementation method, please refer to the relevant content of the above embodiment on determining the target information required for the user interaction request based on the first target session fragment, which will not be repeated here.

[0140] If the relevance score between the currently recalled target session fragment and the user interaction request is less than the set score threshold, then return to execution. Figure 3 Step S13 determines the current content attribute tag to be recalled from multiple content attribute tags based on the recall priority of multiple content attribute tags, and then proceeds with the operations of subsequent steps S14-S16 until the relevance score between the currently recalled target session fragment and the user interaction request is greater than or equal to a set score threshold. Based on the current recalled target session fragment when the relevance score is greater than or equal to the set score threshold, the target information required for the user interaction request is determined. For the specific implementation method, please refer to the relevant content of determining the target information required for the user interaction request based on the first target session fragment in the aforementioned embodiment, which will not be repeated here.

[0141] This embodiment calculates a relevance score between the target session fragment and the user interaction request and sets a threshold, transforming the subjective judgment of whether the recalled information is sufficient to support an answer into an actionable objective standard. If the relevance score of a single recall result to the user interaction request does not reach the set threshold, an automatic completion process is triggered instead of forcibly generating a low-quality response. This feedback loop ensures that the final set of session fragments used to generate the target information meets semantic coverage requirements, effectively preventing problems such as irrelevant answers and insufficient content.

[0142] On the other hand, the cyclical completion process of the recalled information follows the preset recall priority of the content attribute tags. It extracts the second target session fragment from the historical session context corresponding to the highest priority unrecalled tag and combines it with the existing fragments to re-evaluate the relevance between the fragment and the user interaction request. This ensures that knowledge expansion is neither blind full loading nor random splicing, but rather is injected gradually from high to low application value, which can effectively guarantee the order rationality and semantic coherence of cross-session knowledge integration.

[0143] After identifying the target information, it is possible to check whether there is any information in the target information that does not belong to the user boundary identifier preset permissions of the current interaction, based on the user boundary identifier preset permissions of the current interaction. If there is information in the target information that does not belong to the user boundary identifier preset permissions of the current interaction, the information that does not belong to the user boundary identifier preset permissions of the current interaction will be desensitized to obtain the desensitized target information, and the desensitized target information will be output to the user.

[0144] In some embodiments, information regarding permissions preset by the user boundary identifier that does not belong to the current interaction can be encoded to achieve desensitization processing of such information. For example, information regarding permissions preset by the user boundary identifier that does not belong to the current interaction can be replaced with meaningless identifiers, such as "*", "#", etc.

[0145] If the target information does not contain any information with permissions preset by the user boundary identifier that does not belong to the current interaction, the target information is output to the user. This embodiment does not impose permission restrictions during information retrieval, which can ensure the perception capabilities of the digital human associated with the intelligent agent. Permission determination is then performed when the information is subsequently presented to the user, thus acting as a safety barrier and ensuring user data security.

[0146] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 101 and 102 can be device A; or the execution subject of step 101 can be device A, and the execution subject of step 102 can be device B; and so on.

[0147] Furthermore, some processes described in the above embodiments and accompanying drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel.

[0148] Accordingly, embodiments of this application also provide a computer-readable storage medium storing computer instructions, which, when executed by one or more processors, cause one or more processors to perform the steps in the agent-oriented information management method provided in the foregoing embodiments.

[0149] Computer-readable storage media include volatile or non-volatile or a combination thereof, and may be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transfer medium.

[0150] This application also provides a computer program product, including a computer program that, when executed by one or more processors, causes the one or more processors to perform the steps in the agent-oriented information management method provided in the foregoing embodiments.

[0151] In this application embodiment, the specific implementation form of the computer program product is not limited. In some embodiments, the computer program product may be implemented as an application (APP), a mini-program, a PC client, a program module, a plug-in, an installation package, a software development kit (SDK), an image file of an optical disc, a plug-in, or software in the form of Software as a Service (SaaS), etc., but is not limited to these.

[0152] The computer program product should understand that each or a combination of the above-described method flow can be implemented by a computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above-described method embodiments.

[0153] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device includes a memory 40a and a processor 40b. The memory 40a is used to store computer programs and can be configured to store various other data to support operation on a computing platform. Examples of this data include instructions for any application or method used to operate on the electronic device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0154] The processor 40b is coupled to the memory 40a and is used to execute computer programs to perform the steps in the agent-oriented information management method provided in the foregoing embodiments. Specific implementation details of each step can be found in the relevant descriptions of the foregoing embodiments, and will not be repeated here.

[0155] In some alternative implementations, such as Figure 4 As shown, the electronic device may also include optional components such as a communication component 40c, a power supply component 40d, a display component 40e, and an audio component 40f. Figure 4 The diagram only shows some components and does not mean that the electronic device must contain them. Figure 4 The inclusion of all components does not imply that an electronic device can only include... Figure 4 The components shown.

[0156] in addition, Figure 4 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the form factor of the electronic device. The electronic device in this embodiment can be a desktop computer, laptop computer, mobile phone, or IoT device; it can also be a traditional server, cloud server, or server cluster, or other server equipment.

[0157] In this embodiment, the memory is used to store computer programs and can be configured to store various other data to support operation on its host device. The processor can execute the computer programs stored in the memory to implement corresponding control logic. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0158] In the embodiments of this application, the processor can be any hardware processing device capable of executing the above-described method logic. Optionally, the processor can be a central processing unit (CPU), a graphics processing unit (GPU), or a microcontroller unit (MCU); it can also be a programmable device such as a field-programmable gate array (FPGA), a programmable array logic (PAL), a general array logic (GAL), or a complex programmable logic device (CPLD); or it can be an advanced RISC machine (ARM) or a system on chip (SoC), etc., but is not limited thereto.

[0159] In this embodiment, the communication component is configured to facilitate wired or wireless communication between its host device and other devices. The device hosting the communication component can access wireless networks based on communication standards, such as 2G or 3G, 4G, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0160] In embodiments of this application, the display component may include a liquid crystal display (LCD) and a touch panel (TP). If the display component includes a touch panel, the display component can be implemented as a touchscreen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0161] In this embodiment, a power supply component is configured to provide power to various components of the device in which it resides. The power supply component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power supply component resides.

[0162] In embodiments of this application, the audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), which is configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals. For example, in devices with voice interaction capabilities, voice interaction with the user can be achieved through the audio component.

[0163] It should be noted that the terms "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0164] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the aforementioned element.

[0165] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An information management method for intelligent agents, characterized in that, include: Obtain the historical session context generated from the interaction between the target object and the agent in historical sessions; Based on multi-dimensional classification rules, the historical session context is classified in multiple dimensions to determine the first multi-dimensional classification label of the historical session context. The first multi-dimensional classification label is associated with and stored in the historical session context, so that the agent can retrieve the target information required for the user interaction request from the historical session context based on the first multi-dimensional classification label.

2. The method according to claim 1, characterized in that, The step of performing multi-dimensional classification on the historical session context to determine the first multi-dimensional classification label of the historical session context includes: The historical session context and the multi-dimensional classification rules are input into the deep learning model on which the agent depends; The deep learning model is used to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules to obtain the first multi-dimensional classification label.

3. The method according to claim 2, characterized in that, The first multi-dimensional classification label includes a home boundary label; the step of using the deep learning model to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules includes: The deep learning model is used to obtain the classification rules for the attribution boundary labels from the multi-dimensional classification rules; the classification rules for the attribution boundary labels include whether the knowledge content carried by the historical session context has knowledge reusability. Using the deep learning model, entity dependency features, scene binding features, and general knowledge features of the knowledge content are extracted from the historical session context based on the classification rules of the attribution boundary labels. The deep learning model is used to determine whether the knowledge content has knowledge reusability based on the entity dependency features, the scene binding features, and the general knowledge features. If the knowledge content has knowledge reusability, then the ownership boundary label of the historical session context is determined to be a private boundary label. If the knowledge content does not have knowledge reusability, then the attribution boundary label of the historical session context is determined to be a global boundary label.

4. The method according to claim 3, characterized in that, In the case that the attribution boundary label is a private boundary, the first multi-dimensional classification label also includes: task scenario label; The step of using the deep learning model to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules includes: The deep learning model is used to obtain the classification rules for the task scene labels from the multi-dimensional classification rules; the classification rules for the scene labels include classifying application scene types based on the temporal regularity of the target tasks associated with the knowledge content. Using the deep learning model, based on the classification rules of the task scene labels, task semantic features, instruction pattern features, and contextual clue features representing temporal regularity are extracted from the historical session context. Based on the task semantic features, the instruction pattern features, and the context clue features, determine whether the target task has temporal regularity; If the target task exhibits temporal regularity, then the task scenario label of the historical session context is determined to be a steady-state task scenario label; If the target task does not have a temporal regularity, then the task scenario label of the historical session context is determined to be an occasional task scenario label.

5. The method according to claim 3, characterized in that, In the case that the attribution boundary label is a private boundary, the first multi-dimensional classification label further includes: content attribute label; The step of using the deep learning model to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules includes: The deep learning model is used to obtain the classification rules for the content attribute tags from the multi-dimensional classification rules; the classification rules for the content attribute tags include dividing the content attribute tags based on the specific semantics carried by the knowledge content. The deep learning model is used to extract the semantic features of the knowledge content according to the classification rules of the content attribute tags; The deep learning model is used to determine the content attribute tags of the historical session context based on the semantic features and the classification rules of the content attribute tags.

6. The method according to claim 5, characterized in that, The step of using the deep learning model to determine the content attribute tags of the historical session context based on the semantic features includes: If the semantic features indicate that the knowledge content reflects personalized behavioral guidelines, then the content attribute tag is determined to be a preference profile tag; If the semantic features indicate that the knowledge content reflects objective domain cognition, then the content attribute tag is determined to be a factual knowledge tag; If the semantic features indicate that the knowledge content reflects a task execution trajectory that conforms to the application semantics of the agent, then the content attribute tag is determined to be a task execution tag; If the semantic features indicate that the knowledge content reflects the decision-making deduction chain, then the content attribute tag is determined to be a thinking logic tag; If the semantic features of the content indicate that the knowledge content reflects non-task-oriented corpus, then the content attribute tag is determined to be a non-task interaction tag.

7. The method according to claim 3, characterized in that, In the case that the attribution boundary label is a private boundary, the first multi-dimensional classification label further includes: a storage type label; The step of using the deep learning model to perform multi-dimensional classification of the historical session context according to the multi-dimensional classification rules includes: The deep learning model is used to obtain the classification rules for the storage type label from the multi-dimensional classification rules; the classification rules for the storage type label include whether the historical session context contains resource-intensive files; Using the deep learning model, the data types of the data contained in the historical session context are traversed; If the data type of the data contained in the historical session context includes resource-intensive files, then the storage type label of the historical session context is determined to be a compressed storage label; If the data type of the historical session context does not include resource-intensive files, then the storage type label of the historical session context is determined to be the plain text storage label.

8. The method according to claim 1, characterized in that, The step of obtaining the historical session context generated by the interaction between the target object and the agent in historical sessions includes: By using pre-defined hook functions, historical interaction data generated during the interaction between the target object and the agent in a historical session is captured and used as the historical session context.

9. The method according to claim 8, characterized in that, The dimension of the target object is the user dimension; the historical session includes the personal session of the digital human associated with the user identifier corresponding to the target object and the agent, and / or the group session of the digital human with the user identifier and the digital human in the group where the target object is located; the step of capturing historical interaction data generated in the historical session as the historical session context during the interaction between the target object and the agent through a preset hook function includes: The hook function captures the first historical interaction data between the user identifier and the digital human in the personal session, which serves as the historical session context. And / or, The hook function captures second historical interaction data between the user identifier and the digital human in the group session, which serves as the historical session context.

10. The method according to claim 8, characterized in that, The dimension of the target object is a combination of user and group; the historical session includes the group session between the digital human associated with the agent and the user identifier and the digital human in the group where the target object is located; the step of capturing historical interaction data generated in the historical session during the interaction between the target object and the agent through a preset hook function, as the historical session context, includes: The hook function captures third-historical interaction data between the user identifier and the digital human in the group session, serving as the historical session context.

11. The method according to any one of claims 1-10, characterized in that, The agent's historical session context from at least one historical session is associated with and stored with multi-dimensional classification labels; the method further includes asynchronously executing the following steps: Obtain the user interaction request; The user interaction request is subjected to semantic analysis to obtain the interaction intent reflected by the user interaction request; Based on the session identifier corresponding to the user interaction request, determine the user boundary identifier for the current interaction; The user boundary identifier and the interaction intent are mapped to multi-dimensional classification labels to determine the second multi-dimensional classification label corresponding to the user interaction request; From the historical session context of the at least one historical session, the target historical session context associated with the second multi-dimensional classification label is selected; Based on the interaction intent, information retrieval is performed in the target historical session context to obtain the target information.

12. The method according to claim 11, characterized in that, The step of mapping the user boundary identifier and the interaction intent to multi-dimensional classification labels to determine the second multi-dimensional classification label corresponding to the user interaction request includes: The user boundary identifier is converted into the belonging boundary label in the multi-dimensional classification label to obtain the target belonging boundary label; From the interaction intents, extract the scene intent features that represent scene requirements and the content intent features that represent content requirements; The scene intent features are converted into task scene labels in the multi-dimensional classification labels to obtain the target task scene label; The content intent features are converted into content attribute tags in the multi-dimensional classification tags to obtain the target content attribute tags; The target attribution boundary label, the target task scenario label, and the target content attribute label are combined to form the second multi-dimensional classification label.

13. The method according to claim 11, characterized in that, The second multi-dimensional classification label includes: target affiliation boundary label, target task scenario label, and target content attribute label; the step of filtering the target historical session context associated with the second multi-dimensional classification label from the historical session context of the at least one historical session includes: From the historical session contexts of the at least one historical session, select the first historical session context whose affiliation boundary label is the target affiliation boundary label; From the first historical session context, at least a second historical session context with the content attribute tag being the target content attribute tag is selected as a candidate historical session context; The target historical session context is selected from the candidate historical session contexts based on the target task scenario tag.

14. The method according to claim 13, characterized in that, The step of selecting at least one second historical session context from the first historical session context whose content attribute tag is the target content attribute tag, as a candidate historical session context, includes: From the first historical session context, a second historical session context whose content attribute tag is the target content attribute tag is selected as the candidate historical session context; or, If the target content attribute tag is not a preference profile tag, a second historical session context with the target content attribute tag is selected from the first historical session context; and a third historical session context with the preference profile tag is selected from the first historical session context; the second historical session context and the third historical session context are determined as the candidate historical session context.

15. The method according to claim 13, characterized in that, The step of filtering the target historical session context from the candidate historical session contexts based on the target task scenario tag includes: If the target task scenario label is a steady-state task scenario label, then the historical session contexts with the task scenario label being a steady-state task scenario label are selected from the candidate historical session contexts and used as the target historical session context. or, If the target task scenario tag is an occasional task scenario tag, then from the candidate historical session contexts, a fourth historical session context with a steady-state task scenario tag and a fifth historical session context with an occasional task scenario tag are selected; the fourth historical session context and the fifth historical session context are determined as the target historical session context.

16. The method according to claim 11, characterized in that, The step of retrieving information from the target historical session context according to the interaction intent to obtain the target information includes: Semantic feature extraction and keyword feature extraction are performed on the interactive intent to obtain the interactive semantic vector and the first keyword feature; The target historical session context is segmented to obtain multiple candidate session fragments; Semantic feature extraction and keyword feature extraction are performed on the multiple candidate conversation segments respectively to obtain the segment semantic vector and second keyword feature corresponding to each of the multiple candidate conversation segments; Based on the interaction semantic vector and the segment semantic vectors corresponding to the plurality of candidate session segments, a first similarity between the interaction intent and each candidate session segment is determined. Based on the first keyword feature and the second keyword feature corresponding to each of the plurality of candidate conversation segments, a second similarity between the interaction intent and each candidate conversation segment is determined; Based on the first similarity and the second similarity, a first target session segment is determined from the plurality of candidate session segments; The target information is determined based on the first target session fragment.

17. The method according to claim 16, characterized in that, The step of determining the first target session segment from the plurality of candidate session segments based on the first similarity and the second similarity includes: The first similarity and the second similarity are weighted and fused to obtain the baseline recall score for each candidate conversation segment; Obtain the time attribute and historical recall attribute of each of the multiple candidate session segments; For any candidate session segment among the plurality of candidate session segments, the baseline recall score of any candidate session segment is weighted and corrected according to the time attribute and historical recall attribute of the candidate session segment to obtain the target recall score of the candidate session segment. The first target session segment is determined from the plurality of candidate session segments based on their respective target recall scores.

18. The method according to claim 16, characterized in that, There are multiple first target session fragments; these multiple first target session fragments come from the historical session contexts corresponding to multiple content attribute tags; Determining the target information based on the first target session fragment includes: Based on the recall priority of the multiple content attribute tags, the multiple target session fragments are sorted to obtain the target information.

19. The method according to claim 16, characterized in that, Before determining the target information based on the first target session fragment, the method further includes: Calculate the relevance score between the first target session fragment and the user interaction request; If the relevance score is greater than or equal to a set score threshold, then the operation of determining the target information based on the first target session fragment is performed; The method further includes: If the relevance score is less than a set score threshold, the following operations are performed repeatedly until the relevance score between the currently recalled target session fragment and the user interaction request is greater than or equal to the set score threshold; and the target information is determined based on the currently recalled target session fragment; wherein, the operations performed repeatedly include: Based on the recall priority of multiple content attribute tags, determine the content attribute tag to be recalled from the currently unrecalled content attribute tags; In the first historical session context corresponding to the content attribute tag to be recalled, the second target session fragment associated with the interaction intent is recalled; and the first target session fragment and the currently recalled second target session fragment are combined into the target session fragment to be recalled.

20. An information management method for intelligent agents, characterized in that, The system is associated with the historical session context and multi-dimensional classification labels of the historical session context from at least one historical session between the agent and the target object. The method further includes: Obtain user interaction requests; The user interaction request is subjected to semantic analysis to obtain the interaction intent reflected by the user interaction request; Based on the session identifier corresponding to the user interaction request, determine the user boundary identifier for the current interaction; The user boundary identifier and the interaction intent are mapped to multi-dimensional classification labels to determine the second multi-dimensional classification label corresponding to the user interaction request; From the historical session context of the at least one historical session, the target historical session context associated with the second multi-dimensional classification label is selected; Based on the interaction intent, information retrieval is performed in the target historical session context to obtain the target information required for the user interaction request.

21. An electronic device, characterized in that, include: A memory and a processor; wherein the memory is used to store computer programs; The processor is coupled to the memory for executing the computer program to perform the steps of the method according to any one of claims 1-20.

22. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method according to any one of claims 1-20.

23. A computer program product, characterized in that, Includes a computer program that, when executed by one or more processors, causes the one or more processors to perform the steps of the method according to any one of claims 1-20.

Images