Full life cycle AI companion agent system and method

Abstract

The invention relates to the technical field of artificial intelligence interaction, in particular to a full-life-cycle AI accompanying agent system and method.The full-life-cycle AI accompanying agent system comprises a core engine layer, a business service layer, a basic tool layer, a data layer and an interface layer, and the full-life-cycle AI accompanying agent method comprises the steps that multi-modal data, including biological signals, behavior data, voice data, expression data and environment data, of a user are collected; constructing a user digital twinborn model, and updating model parameters through a weighted learning algorithm; identifying a user cognition development stage, and dynamically adjusting interaction complexity and an expression mode; emotional connection is established, and resonance feedback is generated through multi-mode emotional recognition and memory retrieval; according to the method, interaction memory is stored, cross-stage retrieval is realized, data management is performed by adopting a hierarchical memory architecture, and accurate judgment of a cognitive stage is realized by synchronously collecting multi-dimensional data such as behavior tracks, language texts and biological signals of a user and combining a development psychology theoretical model to construct a feature association map.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence interaction technology, and more specifically, to a full-lifecycle AI companion intelligent agent system and method. Background Art

[0002] With the rapid development and widespread application of artificial intelligence (AI) technology, intelligent companion systems have been widely used in a variety of scenarios, including education, healthcare, and elderly care. Existing intelligent companion systems suffer from the following significant technical issues: Existing technologies are generally designed for specific age groups or scenarios. This fragmented design forces users to switch between different AI systems at different life stages, failing to provide a consistent companion experience. Based on short-term interaction data, they lack long-term memory management mechanisms. The effective memory retention period of existing AI systems generally does not exceed six months. Furthermore, they lack an importance grading mechanism, making it impossible to preserve and retrieve memories over decades of time, leading to a fragmented long-term user experience. Psychologist Piaget's theory of cognitive development establishes the stages of human cognitive development, from the sensorimotor stage to the formal operational stage. However, existing AI systems employ fixed interaction modes and expressions, failing to automatically adjust interaction complexity and abstraction based on the user's cognitive development stage, resulting in a sharp decline in applicability during adolescence. With the development of the Internet of Things (IoT), smart device form factors are becoming increasingly diverse. While existing technologies achieve basic device connectivity, they suffer from serious deficiencies in personality consistency, interaction coherence, and user model synchronization. When users switch between different terminals, they need to repeatedly establish context and preferences, which results in a fragmented experience. Existing affective computing technology mainly focuses on emotion recognition and simple responses, lacks a deep emotional connection mechanism, and cannot establish emotional bonds beyond "shallow social connections", especially in long-term companionship scenarios, and cannot provide real emotional support. Summary of the Invention

[0003] To solve the above problems, the purpose of the present invention is to provide a full-life cycle AI companion intelligent agent system and method, and to solve core technical problems such as full-life cycle design, memory continuity, cognitive adaptation, multi-modal consistency and emotional connection through innovative technical means.

[0004] The present invention provides the following technical solutions: In the first aspect, the present invention proposes a full-lifecycle AI companion agent system, specifically including:

[0005] The core engine layer includes:

[0006] Digital twin models for dynamic mapping and prediction of user characteristics;

[0007] A cognitive development manager that identifies the user's cognitive development stage and dynamically adjusts the AI ​​system's expression complexity, level of abstract thinking, and interaction methods;

[0008] An emotional connection system for establishing lasting emotional connections between users and AI through multimodal emotion recognition, emotional memory chaining, and emotional bonding mechanisms;

[0009] A memory map manager for implementing user memory storage, retrieval, and reconstruction; and

[0010] A system scheduler for coordinating the work of the digital twin model, cognitive development manager, emotional connection system, and memory map manager to optimize resource allocation;

[0011] Business service layer, including: prenatal education service module, children's education service module, youth learning service module, adult companionship service module and elderly health care module;

[0012] Used to provide services to users at different life stages;

[0013] Basic tool layer, including:

[0014] Large language models for natural language understanding and generation;

[0015] Affective computing modules for providing emotional understanding and expression capabilities;

[0016] Knowledge graph system for providing structured knowledge support;

[0017] A multimodal understanding module for providing perception and comprehension capabilities;

[0018] Security framework and monitoring and diagnostic modules to support safe and reliable system operation;

[0019] The data layer is responsible for data storage and management, including: user digital twin model data storage and system configuration parameter management;

[0020] The interface layer, including the device adaptation interface, service API, and management interface, is used to achieve connections with external devices and systems.

[0021] As a technical solution of the present invention, the digital twin model is implemented by the following steps:

[0022] Step S1: Obtain multi-dimensional data of the user's entire life cycle and build the original data foundation;

[0023] Step S2, extracting high-dimensional feature vectors from the original data, performing weighted fusion, and generating a standardized user state representation vector;

[0024] Step S3: construct a multimodal dynamic mapping network to achieve real-time synchronous mapping between user status and digital twin model;

[0025] Step S4: Build a prediction model to predict the future evolution trend of user characteristics based on historical user status.

[0026] In step S1, four types of original data, including user biometrics, behavioral patterns, language expressions, and emotional responses, are collected through wearable devices, environmental sensors, and interactive terminals, and stored in a distributed database aligned by timestamps to form a multimodal original data set covering perception, cognition, and emotion.

[0027] As a technical solution of the present invention, the specific implementation of step S2 includes:

[0028] Sub-modal feature extraction, performing feature extraction on each modal data in the multi-modal original data set;

[0029] Weighted fusion and output, using the attention mechanism to calculate the weight of each modality and perform feature splicing to generate multi-dimensional fusion features. After normalization, the representation vector of the user's comprehensive state is generated, denoted as S t .

[0030] As a technical solution of the present invention, the construction of the multimodal dynamic mapping network in step S3 specifically includes:

[0031] The representation vector S of the user's comprehensive state t As input, a bidirectional LSTM is used to capture temporal dependencies and output the initial state of the digital twin model;

[0032] Define a synchronization loss function and iteratively train it through the optimizer to ensure that the error between the digital twin state and the user's actual state is less than a preset value;

[0033] Real-time collection of the user's current new state representation vector, denoted as S t+1 , and cyclically calculate the error with the digital twin state. If the error is greater than the preset value, incremental training is triggered to ensure the accuracy of synchronous mapping.

[0034] As a technical solution of the present invention, the construction of the prediction model in step S4 specifically includes:

[0035] Construct time series features, concatenate the user's past state feature vector set with the real-time state feature vector to form a long time series input sequence;

[0036] Development stage perception, using Piaget's cognitive development stage judgment rules, designing a stage feature weight matrix to adjust the importance of features in different stages;

[0037] The future prediction output uses the Transformer-XL model to process the long time series input sequence to output and cover the user's multi-dimensional future state feature evolution prediction vector, denoted as S t+T ;

[0038] According to the user's real-time state representation vector S t+1 With the feature evolution prediction vector S t+T Generate personalized interaction strategies, collect user interaction feedback information, and update the original dataset.

[0039] As a technical solution of the present invention, the cognitive development manager includes:

[0040] Based on user multimodal data, determine the user's current cognitive development stage, obtain user stage characteristics and user's current cognitive ability;

[0041] Quantify the core indicators of user cognitive ability and determine user development goals;

[0042] Generate adaptive cognitive training strategies based on user stage characteristics and user development goals;

[0043] Establish discrimination rules for each stage of user development, and have experts label the judgment results of each stage of users to form historical labeling data as training data;

[0044] The representation vector S of the user's comprehensive state receiving the digital twin model t , extract sub-features related to cognition to form a cognitive feature vector; take the cognitive feature vector as input and output the probability of each stage to generate the current cognitive development stage label and confidence.

[0045] As a technical solution of the present invention, the emotional connection system includes:

[0046] Multimodal emotion recognizer, used to fuse user multimodal data and identify user emotional state;

[0047] Emotional memory chain, used to create and maintain memory links related to the user's emotional state;

[0048] The emotion generator is used to generate emotional feedback responses based on the user's emotional state and historical memory.

[0049] As a technical solution of the present invention, the memory mapping manager is used to convert the user's discrete memory fragments into structured memory representations and establish a dynamic mapping relationship between memory and current state and future behavior, which will include:

[0050] Receive multi-source raw data from users of the digital twin model and filter out invalid data. Convert the filtered memory fragments into memory tuples in a unified format and output a structured memory database. Use a multimodal Transformer encoder to align memory content of different modalities. Use a bidirectional LSTM network to capture the temporal dependency of memory. Combined with contextual labels, generate a scenario-based memory summary and output a scenario-based memory representation vector.

[0051] Establish an association between historical memory and current user status;

[0052] Adjust the memory representation according to the user's current behavior, eliminate outdated memory, and output an updated database.

[0053] As a technical solution of the present invention, the data layer includes the construction of a short-term memory library, a long-term memory library, and a core memory library, wherein: the short-term memory library adopts vector database technology to store recent interactive memories; the long-term memory library adopts graph database technology to store important historical memories and establish an associated network; the core memory library adopts dedicated time series database technology to permanently store key life events and memories;

[0054] User model data storage management, used to store and record user characteristics and model data, including storing user multi-dimensional feature vectors, recording user development stage history and prediction data, storing user preferences and settings information, and storing user feedback and evaluation data;

[0055] System configuration data management is used to store the configuration information required for the operation of the intelligent system, including storing the configuration parameters of each module, recording system operation logs, storing algorithm-related parameters, and recording system status information.

[0056] The second aspect of the present invention proposes a method for AI accompanying an intelligent agent throughout its life cycle, comprising the following steps:

[0057] Step S100, collecting user multimodal data, including biological signals, behavioral data, voice data, expression data and environmental data;

[0058] Step S200: construct a user digital twin model and update the model parameters through a weighted learning algorithm;

[0059] Step S300: Identify the user's cognitive development stage and dynamically adjust the interaction complexity and expression mode;

[0060] Step S400: establishing an emotional connection and generating resonance feedback through multimodal emotion recognition and memory retrieval;

[0061] Step S500: store interactive memory and implement cross-stage retrieval, using a hierarchical memory architecture for data management.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] This invention simultaneously collects multi-dimensional data such as user behavior, speech, and biometrics, and combines it with theoretical models from developmental psychology to construct a feature association map. By calibrating the combined model using a random forest classifier and expert rules, it achieves accurate determination of cognitive stages. The specific beneficial technical effects are further elaborated in the following detailed embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a system block diagram of the present invention;

[0065] Figure 2 This is a data flow chart of the present invention;

[0066] Figure 3 This is a flowchart of hierarchical memory management in one embodiment of the present invention;

[0067] Figure 4 A flowchart for building a digital twin model in one embodiment of the present invention;

[0068] Figure 5 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0070] The present invention collects multimodal data of users, including biosignal data, behavioral data, voice data, expression data and environmental data; constructs a user digital twin model by extracting user feature vectors, applying stage weights, updating the digital twin model, predicting development trajectories and adaptive fine-tuning; identifies the user's cognitive development stage, determines the cognitive development stage based on the user's age and interaction history, and calculates language complexity parameters and abstract thinking level; selects appropriate companionship strategies according to the user's development stage, including obtaining basic strategies, cognitive adjustments, personalized adjustments and strategy optimization; establishes emotional bonds through emotional connection technology, which is achieved through emotion recognition, emotional memory retrieval, bond strength calculation, emotional resonance generation and emotional memory updating; stores interactive memories and realizes cross-stage memory retrieval, which is achieved through memory importance assessment, memory stratification, association network establishment, context-triggered retrieval and memory reconstruction.

[0071] Collecting multimodal user data: Different types of data are collected at different life stages, including during the fetal period: collecting fetal movement, heart rate and other data through smart belly belts; during childhood: collecting voice, expression and behavior data through smart toys and monitoring equipment; during adolescence: collecting interactive behavior and learning data through learning equipment and mobile phones; during adulthood: collecting living habits, work status and emotional changes through wearable devices and smart homes; during old age: collecting health data and activity patterns through health monitoring equipment and smart homes.

[0072] refer to Figure 1 The embodiment of the present invention discloses a full-lifecycle AI companion agent system comprising:

[0073] The user interaction layer serves as the entry point for direct interaction between the system and users (such as children, parents, teachers, and medical personnel), enabling information input and feedback output. Its design must closely integrate user role requirements, usage scenarios, and cognitive characteristics, aiming to provide an easy-to-use, immersive, and personalized interactive experience.

[0074] The core engine layer is a digital twin model used to realize dynamic mapping and prediction of user characteristics; a cognitive development manager that identifies the user's cognitive development stage and dynamically adjusts the expression complexity, abstract thinking level and interaction mode of the AI ​​system; an emotional connection system that establishes a lasting emotional connection between the user and AI through multimodal emotion recognition, emotional memory chain and emotional bonding mechanism, and realizes the memory mapping manager for user memory storage, retrieval and reconstruction; and a system scheduler that coordinates the work of the digital twin model, cognitive development manager, emotional connection system and memory mapping manager to optimize resource allocation.

[0075] refer to Figure 2 、 Figure 4 , the construction method of the digital twin model is as follows:

[0076] Step S1: Obtain multi-dimensional data of the user's entire life cycle and build the original data foundation.

[0077] Specifically, four types of original data are collected through wearable devices (biometric sensors, smart bracelets), environmental sensors (cameras, millimeter-wave radars), and interactive terminals (microphones, text input devices): user biometrics (heart rate / HRV, EDA, EEG), behavioral patterns (activity trajectories, action sequences), language expressions (voice signals, text content), and emotional responses (facial expressions, voice rhythm). These data are aligned by timestamps and stored in a distributed database to form a multimodal original data set covering perception, cognition, and emotion.

[0078] Step S2: extract high-dimensional feature vectors from the original data, perform weighted fusion, and generate a standardized user state representation vector. Specifically, it includes:

[0079] The sub-modal feature extraction is to extract features from each modal data in the multi-modal original data set.

[0080] More specifically, biological features are extracted through wavelet transform + nonlinear analysis to extract time-frequency domain features (such as SDNN and LF / HF ratio of HRV); behavioral patterns are extracted through DBSCAN clustering + Markov chain; language expression is extracted through MFCC + BERT embedding to extract semantic and rhythmic features; emotional response is extracted through FER2013-CNN + physiological signal association to extract expression / EDA related features.

[0081] Weighted fusion and output, using the attention mechanism to calculate the weight of each modality and perform feature splicing to generate multi-dimensional fusion features. After normalization, the representation vector of the user's comprehensive state is generated, denoted as S t (Represents the user's current physiological, psychological and behavioral comprehensive state).

[0082] Step S3: construct a multimodal dynamic mapping network to achieve real-time synchronous mapping between user status and digital twin model.

[0083] In this step, the construction of the multimodal dynamic mapping network specifically includes: using the representation vector S of the user's comprehensive state t As input, a bidirectional LSTM is used to capture the timing dependency and output the initial state of the digital twin model; a synchronization loss function is defined and iterative training is performed through the optimizer to make the error between the digital twin state and the user's actual state less than the preset value; the representation vector of the user's current new state is collected in real time, denoted as S t+1 , and cyclically calculate the error with the digital twin state. If the error is greater than the preset value, incremental training is triggered to ensure the accuracy of synchronous mapping.

[0084] Step S4: Build a prediction model to predict the future evolution trend of user characteristics based on historical user status.

[0085] In this step, the construction of the prediction model specifically includes: constructing time series features, splicing the user's past state feature vector set with the real-time state feature vector to form a long time series input sequence; development stage perception, using Piaget's cognitive development stage judgment rules, designing a stage feature weight matrix, and adjusting the importance of features in different stages; future prediction output, using the Transformer-XL model to process the long time series input sequence to output and cover the user's multi-dimensional future state feature evolution prediction vector, denoted as S t+T ; According to the user's real-time status representation vector S t+1 With the feature evolution prediction vector S t+T Generate personalized interaction strategies, collect user interaction feedback information, and update the original dataset.

[0086] For example, the user's monthly standardized state vectors over the past year are concatenated with the real-time state to form a 180-day long time series input sequence; Piaget's cognitive development stage judgment rules are introduced (such as "abstract thinking score ≥ 85 points triggers the formal operational period"), the stage weight matrix W is designed, and the importance of features in different stages is adjusted (such as the formal operational period enhances the feature weight of semantic complexity); Transformer-XL (context window 180 days) is used to process the long time series input sequence, and output feature evolution prediction vectors for the next 6 months / 1 year / 3 years, covering multiple dimensions such as cognition (logical reasoning score), emotion (emotional stability index), and sociality (frequency of interpersonal interaction).

[0087] The above embodiments also involve digital twin model verification and robustness optimization: offline verification, using historical 3-month data to calculate synchronization error and prediction accuracy; online optimization, aggregating multi-user data through federated learning (only uploading model gradients to protect privacy), and dynamically adjusting network parameters; introducing Monte Carlo Dropout to generate prediction confidence intervals; edge deployment optimization, lightweighting the model and deploying it to terminal devices such as smart watches to ensure end-to-end latency and support real-time interaction. Closed-loop application and continuous evolution of digital twin models, real-time interaction support, integrating the optimized twin model into the AI ​​companion system, and based on the real-time status S t+1 With the forecast trend S t+T Generate personalized interaction strategies (such as predicting that the user will enter a period of low mood and triggering encouraging conversations in advance); iterate the data closed loop, collect user interaction feedback (such as conversation satisfaction scores), update the original data set D, and update the model parameters, forming a full life cycle closed loop of "data collection, feature fusion, mapping synchronization, and prediction optimization."

[0088] In the present invention, the cognitive development manager is the core component of the digital twin model that focuses on the evolution and intervention of user cognitive abilities. Its technical design deeply integrates developmental psychology theory and artificial intelligence algorithms, aiming to achieve accurate identification of user cognitive stages, ability assessment, and personalized development guidance. Based on user multimodal data, the user's current cognitive development stage is determined, and the user's stage characteristics and current cognitive abilities are obtained; the core indicators of the user's cognitive abilities are quantified to determine the user's development goals; based on the user's stage characteristics and user development goals, an adaptive cognitive training strategy is generated; discrimination rules are established for each stage of the user's development, and the judgment results of each stage of the user are annotated by experts to form historical annotated data as training data; the representation vector S of the user's comprehensive state of the digital twin model is received t , extract sub-features related to cognition to form a cognitive feature vector; take the cognitive feature vector as input and output the probability of each stage to generate the current cognitive development stage label and confidence.

[0089] Stage identification: Based on the user's multimodal data (such as logical reasoning, abstract thinking, and memory ability performance), determine their current cognitive development stage (such as the sensorimotor stage → preoperational stage → concrete operational stage → formal operational stage in Piaget's theory); ability assessment: quantify the core indicators of the user's cognitive ability (such as inductive reasoning scores, analogical thinking accuracy, and working memory capacity); intervention guidance: generate adaptive cognitive training strategies based on stage characteristics and development goals (such as concrete teaching for users in the preoperational stage and abstract problem challenges for users in the formal operational stage).

[0090] It includes a development stage determination module, the purpose of which is to map user behavior performance to the theoretical stages of developmental psychology and solve the problem of "which cognitive stage the user is currently in".

[0091] Integrate Piaget's four-stage theory of cognitive development, sensorimotor period (0-2 years old), preoperational period (2-7 years old), concrete operational period (7-11 years old), formal operational period (11 years old+), and Erikson's psychosocial development stage, such as diligence and inferiority period (6-12 years old), and define the core discrimination rules for each stage, including key indicators of the formal operational period: hypothesis verification ability, propositional thinking, and abstract concept understanding; receive the standardized state vector from the digital twin model, and extract cognition-related sub-features, such as logical reasoning score (classification accuracy from the semantic analysis module), abstract thinking index (based on the semantic complexity of language expression, such as the frequency of metaphor use), working memory capacity, and problem-solving strategies, such as whether to adopt trial and error method → ​​hypothesis verification method. A combined model of random forest classifiers and expert rule calibration is used. Specifically, the random forest classifier is trained based on historically labeled data (experts' judgments on user stages), inputs cognitive feature vectors, and outputs probabilities for each stage, such as an 85% probability for the formal operational stage. Expert rule calibration sets thresholds, such as "Abstract Thinking Index ≥ 70 and Hypothesis Verification Success ≥ 3," to trigger the formal operational stage determination, resolving algorithmic probability ambiguity. The classifier then outputs a label for the current cognitive development stage, such as "Mid-Formal Operational Stage," along with a confidence level, such as 92%.

[0092] It also includes a cognitive ability assessment module, which aims to quantify the core dimensions of the user's cognitive ability and solve the problem of "how strong is the user's current cognitive ability?"

[0093] Based on developmental psychology theory, a three-tiered evaluation index is defined. A Bayesian network model is used to integrate multi-source data, including behavioral task results, language expression semantics, and interactive feedback scores, to calculate the posterior probability distribution of each ability indicator. This outputs a three-dimensional radar chart of the user's cognitive abilities, including logical reasoning, abstract thinking, and working memory, along with percentile rankings for each indicator, such as "abstract thinking ability exceeds 85% of peers."

[0094] In addition, the development intervention strategy generation module aims to generate personalized cognitive training strategies based on stage judgment and ability assessment results to solve the problem of "how to guide users to develop to a higher stage."

[0095] The strategy library is constructed based on the "concrete teaching method" for pre-operational users and the "hypothesis-verification experimental method" for formal operational users, with pre-defined templates for various strategies.

[0096] The system then optimizes policy parameters based on user interaction feedback (such as task completion and satisfaction ratings). These parameters include state input, current stage labels, ability assessment results, and historical policy performance; action space, optional policy templates, and parameters (such as game difficulty and task duration); and reward functions, which combine the user's cognitive ability improvement with interaction engagement or task completion rate. The system then outputs a personalized intervention strategy, such as "Today's training: 30 minutes of graphic reasoning game (level 3 difficulty) + 10 minutes of metaphorical dialogue practice."

[0097] The above data are only examples and do not limit the present invention in any way.

[0098]

[0099] As the core component of the digital twin model, it works in conjunction with other modules through the following table:

[0100] The above recognition algorithm logic and parameter settings in the cognitive stage are shown in the following table:

[0101] stage Judgment indicators Threshold parameter sensorimotor phase Target tracking accuracy <75% <![CDATA[τ1=0.75]]> Preoperational period Conservation concept test score < 60% <![CDATA[τ2=0.60]]> Specific operation period Formal operation test accuracy rate ≥ 80% <![CDATA[τ3=0.80]]>

[0102] The language complexity adjustment formula is as follows for example only:

[0103] LCP=α(0.4V / Vmax+0.3C+0.2A+0.1S)

[0104] In the above formula, α is the coefficient, V is the vocabulary index; C is the syntactic complexity; A is the proportion of abstract concepts; and S is the context adaptability.

[0105] The emotional connection system includes: a multimodal emotion recognizer for fusing user multimodal data and identifying the user's emotional state; an emotional memory chain for creating and maintaining memory links related to the user's emotional state; and an emotion generator for generating emotional feedback responses based on the user's emotional state and historical memory.

[0106] The memory mapping manager is used to convert the user's discrete memory fragments into structured memory representations and establish a dynamic mapping relationship between memory and current status and future behavior. It will include: receiving multi-source original data of users from the digital twin model and filtering invalid data, converting the filtered memory fragments into memory tuples in a unified format, and outputting a structured memory database, using a multimodal Transformer encoder to align memory content of different modalities, using a bidirectional LSTM network to capture the time dependency of memory, combining context labels to generate a scenario-based memory summary, and outputting a scenario-based memory representation vector; establishing an association relationship between historical memory and the current user status; adjusting the memory representation according to the user's current behavior, eliminating outdated memory, and outputting an updated database.

[0107] Specifically:

[0108] Step A: Extract valid memory fragments from the original multimodal data to solve the problem of "which data belongs to memory".

[0109] Implementation method:

[0110] Receive multi-source raw data from the digital twin model, classified by modality as shown in the following table:

[0111]

[0112] A sliding window combined with key event detection is used to filter invalid data. Specifically, continuous data is segmented by time granularity, retaining high-frequency features within the window; key memory points are marked using a rule engine or machine learning model; the filtered memory fragments are converted into memory tuples in a unified format (t, m, c); and a structured memory database (a collection of memory tuples sorted by timestamp) is output. Where t is the timestamp, which can be accurate to the second; m is the memory content, which can be an embedding vector of behavior / language / biological signals, such as a graph neural network embedding of a behavior trajectory; and c is a context label, such as "work scene," "home scene," or "learning scene."

[0113] Step B: Combine the discrete memory elements into a high-dimensional memory representation vector to solve the problem of "how to describe user memory with a unified vector". The implementation method is as follows:

[0114] A multimodal Transformer encoder is used to align memory content across different modalities. A bidirectional LSTM network is used to capture the temporal dependencies of memories. Combined with the contextual label c, a conditional variational autoencoder is used to generate a contextualized memory summary. The output is a contextualized memory representation vector (integrating multidimensional memory information across time, modality, and context).

[0115] Step C: Establish the association between historical memory and the current user status to solve the problem of "how historical memory affects the current status". The implementation method is as follows:

[0116] Based on the current state (normalized vectors from the digital twin model), relevant memory fragments are recalled from the memory database through similarity retrieval (e.g., cosine similarity). A gated recurrent unit is used to calculate the contribution of historical memory to the current state. Counterfactual reasoning is used to verify the significance of the memory's impact on the state. The output is a memory influence factor (quantifying the contribution of historical memory to the current state) and a list of associated memories.

[0117] Step D adjusts the memory representation based on the user's current behavior and eliminates outdated memories, solving the problem of "how to maintain the effectiveness of memory." The implementation method is as follows:

[0118] For high-impact memory fragments, gradient backpropagation is used to increase their weight in the memory database. An exponential decay model is used to automatically reduce the weight of low-impact memories. Users are allowed to actively tag sensitive memories (such as "medical records") and differential privacy techniques are used to add noise to related memory tuples to ensure data is untraceable. The updated memory database (a collection of memory tuples with adjusted weights) is then output.

[0119] In one embodiment, the business service layer of the system includes: prenatal education service module, children's education service module, youth learning service module, adult companionship service module and elderly health care module; which are used to provide services for users at different life stages.

[0120] For example, the prenatal education service module includes a fetal movement monitoring and analysis system, a personalized prenatal education content generator, a maternal status monitoring system and an early interaction recorder; the children's education service module includes a development stage adaptation system, an interest exploration engine, a habit formation system, a safety monitoring network and an enlightenment education content library; the youth learning service module includes a subject knowledge map, a thinking cultivation system, an emotion management counselor, a social skills cultivation system and a career exploration platform; the adult companionship service module includes a stress management system, a career development consultant, an emotional support network, a life management assistant and a self-growth promoter; the elderly health care module includes a health status monitoring system, a cognitive function maintainer, a memory reproduction system, a daily life assistance network and a social connection maintainer.

[0121] Exemplary:

[0122] Fetal companionship stage

[0123] A smart belly belt collects fetal movement signals, and the fetal status is analyzed using wavelet transform filtering and an LSTM model. A prenatal education content generation algorithm combines parental preferences and fetal developmental stage to generate personalized music and voice stimulation content. The system records the fetus's response patterns to different stimuli and establishes a preliminary interactive memory library.

[0124] Children's education stage

[0125] Multimodal data is collected through smart toys and children's watches, and a cognitive stage recognition algorithm is used to determine a child's developmental stage. A content adaptation system dynamically adjusts language complexity and interaction patterns, while a safety monitoring module monitors environmental risks in real time. A behavioral guidance algorithm combines positive reinforcement and gamification strategies to promote healthy development.

[0126] Adolescent learning stage

[0127] Build a subject knowledge map and design personalized learning paths based on Vygotsky's theory. The emotional companionship system identifies adolescent mood swings, providing reflective guidance and solution exploration. The social assistance module uses scenario simulations to train social skills and promote adolescents' social adaptability.

[0128] Hierarchical memory management (see Figure 3 )

[0129] Adopting a three-level memory storage architecture:

[0130] Short-term memory library (Milvus): stores interaction data within 7 to 30 days; long-term memory library (Neo4j): establishes an event-emotion-person association network; core memory library (InfluxDB): permanently stores major life events; memory retrieval uses multi-feature similarity calculation, for example, the context similarity weight accounts for 60% and the emotion similarity accounts for 30%.

[0131] The basic tool layer includes a large language model for natural language understanding and generation capabilities, an emotional computing module for emotional understanding and expression, a knowledge graph system for structured knowledge support, a multimodal understanding module for perception and comprehension capabilities, and a security framework and monitoring and diagnosis module to support safe and reliable system operation. This provides support for basic AI functions.

[0132] refer to Figure 4 The data layer is responsible for data storage and management, including user digital twin model data storage and system configuration parameter management. The data layer is the underlying infrastructure of the digital twin system in this invention. It is responsible for accessing, storing, governing, and managing the entire service lifecycle of multi-source heterogeneous data, providing highly reliable, low-latency, and scalable data support for upper-layer memory map managers, cognitive training modules, and emotional intervention modules.

[0133] In this embodiment, the data layer includes the construction of a short-term memory library, a long-term memory library, and a core memory library, wherein: the short-term memory library adopts vector database technology to store recent interactive memories; the long-term memory library adopts graph database technology to store important historical memories and establish an associated network; the core memory library adopts a dedicated time series database technology to permanently store key life events and memories; user model data storage management is used to store and record user characteristics and model data, including storing user multi-dimensional feature vectors, recording user development stage history and prediction data, storing user preferences and settings information, and storing user feedback and evaluation data; system configuration data management is used to store the configuration information required for the operation of the intelligent system, including storing the configuration parameters of each module, recording the system operation log, storing algorithm-related parameters, and recording system status information. The hierarchical management of core memory, long-term memory, and short-term memory refers to Figure 3 .

[0134] The system of the present invention also includes an interface layer, including a device adaptation interface, a service API and a management interface, which are used to achieve connection with external devices and systems.

[0135] The interface layer is the core hub component of the digital twin cognitive development management system in this invention, fulfilling the key responsibilities of "connecting internal modules, docking with external systems, and standardizing data interaction." Its design goal is to address issues such as communication protocol heterogeneity, data format incompatibility, and security risk dispersion between the system's multiple internal modules (data layer, cognitive development manager, user interaction layer) and the external ecosystem (third-party services, hardware devices), ensuring the efficient, stable, and secure collaborative operation of the entire system.

[0136] As an internal communication bridge, it connects the data layer (storage and governance) and the cognitive development manager (assessment and intervention), as well as the cognitive development manager and the user interaction layer (input and output), to achieve standardization of data flow between multiple modules (such as unified data format, transmission protocol) and normalization of calling processes. As an external ecological adapter, it connects to external hardware devices (such as sensors, wearable devices), third-party systems (such as education platforms, medical information systems) and user terminals (such as APP, PC), and solves the private protocol barriers of devices from different manufacturers (such as manufacturer A's sensors use a private binary protocol, and manufacturer B uses MQTT), as well as the data format differences between different systems (such as XML, JSON, Protobuf). As the "first line of defense" between the system and the outside world, it prevents illegal access (such as unauthorized third-party calls), malicious attacks (such as DDoS) and data leakage (such as clear text transmission of sensitive information) through mechanisms such as interface authentication, traffic filtering, and protocol verification.

[0137] The interface layer is a common means of general AI systems. Its specific design scheme refers to the existing system for the design of various interface modules (such as parameter settings, protocol settings, data conversion methods, message push methods and system data security, etc.), which will not be elaborated here.

[0138] The second aspect of the present invention proposes a method for AI companionship of an intelligent body throughout its life cycle, comprising the following steps: Step S100, collecting multimodal data of the user, including biological signals, behavioral data, voice data, expression data and environmental data; Step S200, constructing a digital twin model of the user, and updating the model parameters through a weighted learning algorithm; Step S300, identifying the user's cognitive development stage, and dynamically adjusting the complexity and expression of the interaction; Step S400, establishing an emotional connection, and generating resonant feedback through multimodal emotion recognition and memory retrieval; Step S500, storing interactive memory and realizing cross-stage retrieval, and adopting a hierarchical memory architecture for data management.

[0139] The method has been specifically described in the above-mentioned system embodiments and will not be described in detail here.

[0140] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same.

Claims

1. The full life cycle AI companion intelligent agent system is characterized by: include: The core engine layer includes: Digital twin models for dynamic mapping and prediction of user characteristics; A cognitive development manager that identifies the user's cognitive development stage and dynamically adjusts the AI ​​system's expression complexity, level of abstract thinking, and interaction methods; An emotional connection system for establishing lasting emotional connections between users and AI through multimodal emotion recognition, emotional memory chaining, and emotional bonding mechanisms; A memory map manager for implementing user memory storage, retrieval, and reconstruction; and A system scheduler for coordinating the work of the digital twin model, cognitive development manager, emotional connection system, and memory map manager to optimize resource allocation; Business service layer, including: prenatal education service module, children's education service module, youth learning service module, adult companionship service module and elderly health care module; Used to provide services to users at different life stages; Basic tool layer, including: Large language models for natural language understanding and generation; Affective computing modules for providing emotional understanding and expression capabilities; Knowledge graph system for providing structured knowledge support; A multimodal understanding module for providing perception and comprehension capabilities; Security framework and monitoring and diagnostic modules to support safe and reliable system operation; The data layer is responsible for data storage and management, including: user digital twin model data storage and system configuration parameter management; The interface layer, including the device adaptation interface, service API, and management interface, is used to achieve connections with external devices and systems.

2. The full life cycle AI companion agent system according to claim 1 is characterized in that: The digital twin model is implemented by the following steps: Step S1: Obtain multi-dimensional data of the user's entire life cycle and build the original data foundation; Step S2, extracting high-dimensional feature vectors from the original data, performing weighted fusion, and generating a standardized user state representation vector; Step S3: construct a multimodal dynamic mapping network to achieve real-time synchronous mapping between user status and digital twin model; Step S4: Build a prediction model to predict the future evolution trend of user characteristics based on historical user status. In step S1, four types of original data, including user biometrics, behavioral patterns, language expressions, and emotional responses, are collected through wearable devices, environmental sensors, and interactive terminals, and stored in a distributed database aligned by timestamps to form a multimodal original data set covering perception, cognition, and emotion.

3. The full life cycle AI companion agent system according to claim 2 is characterized in that: The specific implementation of step S2 includes: Sub-modal feature extraction, performing feature extraction on each modal data in the multi-modal original data set; Weighted fusion and output, using the attention mechanism to calculate the weight of each modality and perform feature splicing to generate multi-dimensional fusion features. After normalization, the representation vector of the user's comprehensive state is generated, denoted as S t .

4. The full life cycle AI companion agent system according to claim 3 is characterized in that: The construction of the multimodal dynamic mapping network in step S3 specifically includes: The representation vector S of the user's comprehensive state t As input, a bidirectional LSTM is used to capture temporal dependencies and output the initial state of the digital twin model; Define a synchronization loss function and iteratively train it through the optimizer to ensure that the error between the digital twin state and the user's actual state is less than a preset value; Real-time collection of the user's current new state representation vector, denoted as S t+1 , and cyclically calculate the error with the digital twin state. If the error is greater than the preset value, incremental training is triggered to ensure the accuracy of synchronous mapping.

5. The full life cycle AI companion agent system according to claim 4 is characterized in that: The construction of the prediction model in step S4 specifically includes: Construct time series features, concatenate the user's past state feature vector set with the real-time state feature vector to form a long time series input sequence; Development stage perception, using Piaget's cognitive development stage judgment rules, designing a stage feature weight matrix to adjust the importance of features in different stages; The future prediction output uses the Transformer-XL model to process the long time series input sequence to output and cover the user's multi-dimensional future state feature evolution prediction vector, denoted as S t+T ; According to the user's real-time state representation vector S t+1 With the feature evolution prediction vector S t+T Generate personalized interaction strategies, collect user interaction feedback information, and update the original dataset.

6. The full life cycle AI companion agent system according to claim 1 is characterized in that: The cognitive development manager includes: Based on user multimodal data, determine the user's current cognitive development stage, obtain user stage characteristics and user's current cognitive ability; Quantify the core indicators of user cognitive ability and determine user development goals; Generate adaptive cognitive training strategies based on user stage characteristics and user development goals; Establish discrimination rules for each stage of user development, and have experts label the judgment results of each stage of users to form historical labeling data as training data; The representation vector S of the user's comprehensive state receiving the digital twin model t , extract sub-features related to cognition to form a cognitive feature vector; take the cognitive feature vector as input and output the probability of each stage to generate the current cognitive development stage label and confidence.

7. The full life cycle AI companion agent system according to claim 1, characterized in that: The emotional connection system includes: Multimodal emotion recognizer, used to fuse user multimodal data and identify user emotional state; Emotional memory chain, used to create and maintain memory links related to the user's emotional state; The emotion generator is used to generate emotional feedback responses based on the user's emotional state and historical memory.

8. The full life cycle AI companion agent system according to claim 1, characterized in that: The memory mapping manager is used to convert discrete user memory fragments into structured memory representations and establish a dynamic mapping relationship between memory, current state, and future behavior. It will include: Receive multi-source raw data from users of the digital twin model and filter out invalid data. Convert the filtered memory fragments into memory tuples in a unified format and output a structured memory database. Use a multimodal Transformer encoder to align memory content of different modalities. Use a bidirectional LSTM network to capture the temporal dependency of memory. Combined with contextual labels, generate a scenario-based memory summary and output a scenario-based memory representation vector. Establish an association between historical memory and current user status; Adjust the memory representation according to the user's current behavior, eliminate outdated memory, and output an updated database.

9. The full life cycle AI companion agent system according to claim 1, characterized in that: The data layer includes the construction of a short-term memory bank, a long-term memory bank, and a core memory bank. Among them, the short-term memory bank uses vector database technology to store recent interactive memories; the long-term memory bank uses graph database technology to store important historical memories and establish an associated network; the core memory bank uses a dedicated time series database technology to permanently store key life events and memories; User model data storage management, used to store and record user characteristics and model data, including storing user multi-dimensional feature vectors, recording user development stage history and prediction data, storing user preferences and settings information, and storing user feedback and evaluation data; System configuration data management is used to store the configuration information required for the operation of the intelligent system, including storing the configuration parameters of each module, recording system operation logs, storing algorithm-related parameters, and recording system status information.

10. A method for a full life cycle AI companion agent system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Step S100, collecting user multimodal data, including biological signals, behavioral data, voice data, expression data and environmental data; Step S200: construct a user digital twin model and update the model parameters through a weighted learning algorithm; Step S300: Identify the user's cognitive development stage and dynamically adjust the interaction complexity and expression mode; Step S400: establishing an emotional connection and generating resonance feedback through multimodal emotion recognition and memory retrieval; Step S500: store interactive memory and implement cross-stage retrieval, using a hierarchical memory architecture for data management.

Images