Agent cognitive self-evolution method and system based on dynamic memory graph

Abstract

The application relates to the technical field of computer software, and discloses an agent cognitive self-evolution method and system based on a dynamic memory graph. The application constructs a memory graph for each agent; long-term memory and short-term memory are extracted from the memory graph and are respectively converted into prompts; external input is mapped into an instantaneous memory prompt; the three types of prompts are subjected to vector concatenation and dimension alignment in a feature layer to form a joint prompt space, which drives a multi-modal large model to generate a cognitive response; the instantaneous memory prompt is modeled into a probability distribution and is stabilized; cross-task shared knowledge is extracted, and the long-term and short-term memory prompts are updated; the cognitive response is output to a multi-agent simulation environment, a global situation is dynamically updated, and a new round of feedback is generated to form a cognitive self-evolution closed loop. The application realizes continuous learning and knowledge transfer of the agent in a dynamic environment, and significantly improves the cognitive coherence of the multi-agent system in complex task deduction and the explanatory power of group evolution simulation.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and specifically to a method and system for the cognitive self-evolution of intelligent agents based on dynamic memory graphs. Background Technology

[0002] As the basic unit of social network evolution, the individual's cognitive state directly determines the processes of information reception, understanding, memory, judgment, and behavioral expression. Traditional social simulation research mainly focuses on information dissemination paths and behavioral rule design, paying insufficient attention to the continuous construction and dynamic evolution of the individual's internal cognitive structure, making it difficult to explain complex phenomena such as idea transfer, stance shift, and group polarization under long-term information exposure. Therefore, modeling methods based on the self-evolution of individual cognition have gradually gained attention. The core objective is to construct a cognitive system at the agent level with long-term memory, emotional state, value preferences, and reflective abilities, enabling the agent to form a self-consistent and historically consistent cognitive trajectory through continuous interaction, and thereby achieve cross-scale simulation and prediction from micro-individual behavior to macro-social trends.

[0003] Research on the self-evolution of individual cognition aims to characterize the mechanisms of cognitive formation, memory accumulation, and knowledge transfer in individuals during continuous information interaction. It is an important research direction connecting cognitive science, social computing, and artificial intelligence. By endowing intelligent agents with continuous learning and cognitive updating capabilities, it enables them to accumulate knowledge and adapt behavior in dynamic environments, thereby driving the evolution of social simulation from rule-driven to cognition-driven. Its core value is mainly reflected in the following aspects: 1) Dynamic memory map modeling: Drawing on human memory mechanisms, a dynamic memory map system is constructed that integrates long-term memory, short-term memory, and transient memory to achieve structured storage, dynamic retrieval, and continuous updating of cognitive knowledge. 2) Continuous learning and knowledge transfer: Simulating the continuous absorption, integration, and reconstruction of external information by individuals, enabling intelligent agents to continuously learn new knowledge while retaining historical experience, and achieving cross-task and cross-scenario knowledge transfer and environmental adaptation. 3) Dynamic evolution of cognitive states: Establishing an evolutionary mechanism for the continuous change of cognitive states over time, characterizing the dynamic adjustment process of emotions, attitudes, preferences, and value judgments under the influence of information stimuli and social interactions. 4) Cognitive-driven social simulation: Taking individual cognitive evolution as the core driving force, it constructs a feedback loop of "information dissemination - cognitive updating - behavioral decision-making" to realize cross-scale mapping of micro-cognitive changes to macro-group behavioral emergence.

[0004] Traditional methods based on ABMs (Agent-based Models) typically rely on static rules or fixed strategies for individual modeling, making it difficult to cope with the dynamic evolution of information and behavioral heterogeneity in complex social environments. While these methods have achieved some success in closed scenarios, they often lack generalization ability in real-world tasks, exhibiting problems such as delayed cognitive updates and rigid behavioral responses. In recent years, research has begun to explore the introduction of individual modeling mechanisms. One representative approach is to introduce policy adaptation mechanisms such as reinforcement learning and evolutionary game theory, enabling agents to continuously adjust their behavioral strategies based on environmental feedback. Another approach leverages pre-trained language models to empower agents, simulating their language expression, viewpoints, and reasoning judgments, thereby enhancing their expressiveness and response diversity. In dynamic multi-agent environments, efficiently absorbing new task information while retaining long-term prior knowledge is crucial. Secondly, existing memories are mostly isolated text records, making cross-task knowledge mining and reconstruction capabilities important; instantaneous perceptions in new tasks are naturally integrated into the long and short-term memory system. This fragmented, non-dynamically updated memory mechanism enables current agents to achieve continuous knowledge accumulation and cognitive self-evolution in real dynamic environments.

[0005] However, while the aforementioned methods improve the detailed performance of individual simulations, most remain at the adaptive optimization level of the strategy or response layer, lacking continuous construction and endogenous evolutionary modeling of cognitive structures. Specifically, these agents typically cannot form a self-consistent, historically consistent, and personality-coherent cognitive system based on the reception of external information, the integration of internal memories, and the evolution of emotions and preferences in long-term interactions. This deficiency limits the depth of an individual's understanding of events, their ability to shape changes in viewpoints and positions, and the interpretable modeling of the group's co-evolutionary process. First, structural modeling is weak. Existing methods mostly rely on immediate context for response generation, lacking explicit modeling and long-term maintenance of deep cognitive structures such as stable personality traits, value positions, and social identities, making it difficult for agents to maintain consistency in cognition and behavior in complex interactions across situations and events. Second, evolutionary mechanisms are lacking. Most systems only rely on short-term task feedback for strategy updates, lacking the ability to model the dynamic evolution of cognitive states in continuous social interactions, making it difficult to realistically depict the gradual changes in an individual's concepts, attitudes, and preferences under long-term information exposure. Finally, there is the fragmentation of memory mechanisms. Although some models have introduced memory modules, they mostly remain at the level of simple historical records or dialogue caches, lacking structured organization and reasoning mechanisms for multi-layered memory content (such as event experiences, emotional fluctuations, and the evolution of viewpoints). This results in individual behavior exhibiting a fragmented temporal characteristic, making it difficult to reflect a cognitive trajectory with historical continuity and contextual consistency. In conclusion, the continuous construction of cognitive structures and the multi-layered integration of historical memory in the self-evolution of individual cognition are important challenges that urgently need to be overcome in the current modeling of social simulated intelligent agents. Summary of the Invention

[0006] To overcome the core bottlenecks in existing multi-agent simulations, such as shallow individual cognitive modeling, fragmented memory mechanisms, and a lack of knowledge transfer and continuous learning capabilities in dynamic interactive environments, this invention provides a method and system for agent cognitive self-evolution based on a dynamic memory graph. Based on a human-like memory transfer mechanism, this invention innovatively parameterizes the memory modules of individual agents in a large multimodal model, constructing a memory graph encompassing long-term, short-term, and transient memory. A memory controller dynamically embeds memory modules at different levels into the large multimodal model through a prompting optimization mechanism, achieving cross-level knowledge integration and continuous updating. Specifically, this invention extends long-term and short-term memory prompts to include the individual's prior cognition. During memory processing, transient memory prompts replace the inherent input sequence and are integrated with fixed prior memory prompts, thereby guiding the agent to acquire new knowledge while retaining its historical cognitive foundation. To address the "high stability-strong plasticity" dilemma in dynamic evolution, this invention innovatively proposes a Bayesian distribution alignment technique for instantaneous memory cues. Utilizing a Bayesian neural network, it accurately simulates the perceptual distribution of instantaneous memories in different dynamic tasks in an adversarial manner, achieving efficient modeling and storage of immediate experience. Furthermore, this invention constructs a cognitive feedback loop based on cross-task knowledge mining. During the evolution of a new task, stored instantaneous memories not only assist in activating the recall and reconstruction of historical knowledge but also seamlessly feed back and integrate refined instantaneous perceptual experiences into short-term and long-term memories, achieving knowledge integration from the instantaneous state to the long-term state. Through this self-evolutionary mechanism integrating memory parameterization, Bayesian distribution alignment, and cross-layer knowledge integration, this invention endows individual intelligent agents with a human-like instinct for knowledge accumulation and adaptability to dynamic environments, significantly improving the cognitive coherence and explanatory power of multi-agent systems in complex task deduction and group evolution modeling.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for the cognitive self-evolution of an intelligent agent based on a dynamic memory map, comprising the following steps: S1, construct a memory map for each agent that includes long-term memory, short-term memory, and transient memory; S2, select long-term memory nodes and short-term memory nodes from the memory map, and generate long-term memory cues and short-term memory cues respectively through aggregation, encoding and cue mapping; S3 maps external inputs to immediate memory cues; the immediate memory cues are concatenated with long-term memory cues and short-term memory cues at the feature layer and aligned in dimensions to form a joint cue space, which drives the multimodal large model to generate cognitive responses; S4, model the transient memory cue as a probability distribution, and use the Bayesian distribution alignment mechanism to perform parameter constraints and stabilization to obtain a stable transient memory representation; S5 encapsulates stable transient memory representations into historical experience units and stores them in the historical memory pool. Through cross-task retrieval and knowledge mining, cross-task shared knowledge is extracted, and this shared knowledge is used to update short-term memory cues and long-term memory cues. S6, the cognitive response is output to the multi-agent simulation environment, the environment server captures the behavioral output of each agent, dynamically updates the global situation and generates a new round of external input feedback to each agent, and returns to step S3.

[0008] In one embodiment, mapping external input to transient memory prompts specifically includes: Through perceptual coding network Extracting external inputs from multimodal modes The high-dimensional representation is then mapped through a learnable projection layer to a continuous cue space aligned with the input dimension of the multimodal large model, generating transient memory cue. ,in, Indicates the agent index. This represents the index at time t.

[0009] In one embodiment, the step of concatenating and aligning the immediate memory cue with the long-term memory cue and the short-term memory cue at the feature layer to form a joint cue space specifically includes: Through dynamic memory map fusion function Constructing a federated suggestion space ,in, For long-term memory cues, This serves as a short-term memory cue for the previous moment. This serves as a prompt for the current immediate memory.

[0010] In one embodiment, modeling the transient memory cue as a probability distribution and using a Bayesian distribution alignment mechanism for parameterization and stabilization to obtain a stable transient memory representation specifically includes: Current instantaneous memory prompt Modeled as an approximate posterior distribution : ; For the stable representation of immediate memory, The cognitive response generated by the agent based on the current input. It follows a Gaussian distribution. To stabilize the central representation of immediate memory, The uncertainty of current instantaneous memory; Sampling yields a stable instantaneous memory representation : ; in, , For random perturbations sampled from the standard normal distribution, It is the identity matrix. This is element-wise multiplication; Constructing a ternary joint loss function Update the posterior distribution of the immediate memory: ; in, For task deviation loss, To constrain the approximate posterior distribution With historical prior distribution KL divergence term of parameter offset, To counteract training losses, , These are the weighting coefficients. This serves as a reminder for immediate memory from the previous moment; After optimization using the joint loss function, a stable instantaneous memory representation aligned with the Bayesian distribution is obtained. .

[0011] In one embodiment, the adversarial training loss specifically includes: Introducing a discriminator network It distinguishes between instantaneous memory sampling features and target task environmental features through game theory, and combats training loss. for: ; in, The distribution of real-world environmental features for the target task. Expressing expectations, Represents external input samples It conforms to the distribution of real-world environmental characteristics of the target task.

[0012] In one embodiment, task deviation loss The calculation formula is: ; in, Indicates the target response for the current task. This represents the response semantic encoding function. This represents the L2 norm.

[0013] In one embodiment, the process of encapsulating stable transient memory representations into historical experience units and storing them in a historical memory pool, and then extracting cross-task shared knowledge through cross-task retrieval and knowledge mining, specifically includes: Representing stable transient memory and the corresponding task scenarios Environmental conditions and cognitive response They are collectively encapsulated as historical experience units. : ; Units of historical experience Stored in the historical memory pool ,Right now ; By retrieval function From the pool of historical memory Activate the set of candidate experiences that are potentially relevant to the current task. : ; Through knowledge mining functions Semantic induction and structural abstraction are performed on the candidate experience set to extract cross-task shared knowledge. .

[0014] In one embodiment, updating short-term memory cues and long-term memory cues using the shared knowledge specifically includes: Based on cross-task knowledge sharing Frequency of occurrence in historical missions The degree of contribution to the completion of the task and consistency with existing cognitive systems Construct knowledge scoring When the knowledge score exceeds a preset threshold, the cross-task shared knowledge will be... Synchronized updates to short-term memory tips and long-term memory tips : , ; in, These are the short-term memory cues from the previous moment and the current short-term memory cues, respectively. For short-time memory update functions, This is a function for updating long-term memory.

[0015] In one embodiment, knowledge scoring The calculation method is as follows: ; in, , , These are weight parameters; The calculation method is as follows: ; in, Indicates the first An intelligent agent at a historical moment Stored in the historical memory pool Historical experience unit, express Corresponding cross-task knowledge sharing, Represents the similarity function. Indicates the time decay weight; This represents the response semantic encoding function; This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and takes the value 0 if the condition within the parentheses is not met. The calculation method is as follows: ; in To retrieve the relevant subset of historical experiences that have been hit, The preset similarity threshold; This represents the reward value for historical missions. Represents the normalization function; The calculation method is as follows: ; in, Represents cosine similarity. This serves as a short-term memory cue for the previous moment.

[0016] In a second aspect, the present invention provides an intelligent agent cognitive self-evolution system based on a dynamic memory graph, for implementing the method of any embodiment of the first aspect, comprising: The memory map construction module is used to construct a memory map for each agent, which includes long-term memory, short-term memory, and transient memory. The memory cue generation module is used to filter long-term memory nodes and short-term memory nodes from the memory graph, and generate long-term memory cues and short-term memory cues respectively through aggregation, encoding and cue mapping; The joint cognition construction module is used to concatenate and align the vectors of immediate memory cues, long-term memory cues, and short-term memory cues at the feature layer to form a joint cue space, which drives the multimodal large model to generate cognitive responses. The Bayesian distribution alignment module is used to model transient memory cues as a probability distribution, and obtains a stable transient memory representation by parameterizing and stabilizing them through the Bayesian distribution alignment mechanism. The cross-task knowledge integration module is used to encapsulate stable transient memory representations into historical experience units and store them in the historical memory pool. Cross-task shared knowledge is extracted through cross-task retrieval and knowledge mining, and the shared knowledge is used to update short-term memory cues and long-term memory cues. The environment interaction and iterative cycle module is used to output the cognitive response to the multi-agent simulation environment. The environment server captures the behavioral output of each agent, dynamically updates the global situation, generates a new round of external input feedback to each agent, and triggers the joint cognitive construction module to be executed again.

[0017] Compared with the prior art, the beneficial technical effects of the present invention are: The efficiency of continuous learning and self-evolution of individual agent cognition is significantly improved: through a dynamic memory map-based continuous guidance mechanism that integrates long-term memory, contextual memory, and immediate attention, the agent dynamically adjusts its cognitive state in multiple rounds of interaction, autonomously forming a stable yet evolvable cognitive personality. Compared to static labels or rule models, the agent can exhibit human-like experience accumulation, strategy updates, and cognitive shifts in the evolution of public opinion, possessing human-like growth potential and individual consistency.

[0018] Enhanced cognitive consistency and adaptability in complex dynamic environments: Through a closed-loop mechanism of perception-modeling-alignment-mining-updating, the agent can continuously adjust its cognitive state in a dynamically changing environment, achieving adaptive response to changes in context and task distribution, thereby improving its adaptability and stability in complex social system simulation, group behavior evolution modeling, and long-term decision-making scenarios.

[0019] In summary, this invention effectively endows individual intelligent agents with human-like continuous learning capabilities and the ability to adapt to dynamic changes in the public opinion environment in large-scale social simulations, further expanding the application boundaries of artificial intelligence in fields such as individual behavior modeling, cognitive evolution analysis, and social system simulation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, the present invention provides a method for the cognitive self-evolution of an intelligent agent based on a dynamic memory graph, comprising the following steps: S1, construct a memory map for each agent that includes long-term memory, short-term memory, and transient memory; S2, select long-term memory nodes and short-term memory nodes from the memory map, and generate long-term memory cues and short-term memory cues respectively through aggregation, encoding and cue mapping; S3 maps external inputs to immediate memory cues; the immediate memory cues are concatenated with long-term memory cues and short-term memory cues at the feature layer and aligned in dimensions to form a joint cue space, which drives the multimodal large model to generate cognitive responses; S4, model the transient memory cue as a probability distribution, and use the Bayesian distribution alignment mechanism to perform parameter constraints and stabilization to obtain a stable transient memory representation; S5 encapsulates stable transient memory representations into historical experience units and stores them in the historical memory pool. Through cross-task retrieval and knowledge mining, cross-task shared knowledge is extracted, and this shared knowledge is used to update short-term memory cues and long-term memory cues. S6, the cognitive response is output to the multi-agent simulation environment, the environment server captures the behavioral output of each agent, dynamically updates the global situation and generates a new round of external input feedback to each agent, and returns to step S3.

[0023] The following is a breakdown into several parts.

[0024] 1. Environment setup and agent initialization.

[0025] In real social networks, individual heterogeneity is widely manifested across multiple dimensions, including cognitive style, emotional preferences, social roles, and behavioral logic, making it a key driver of information dissemination and public opinion evolution. At the environmental level, the simulation system integrates multi-source heterogeneous social data such as text, images, videos, and interactive behaviors. Through a unified semantic fusion interface and cross-modal alignment mechanism, it provides individuals with realistic, complex, and dynamic social input scenarios. This environment not only supports the injection of multimodal data streams but also simulates social relationship network structures, role identity distribution, and interactive feedback mechanisms, forming a multi-dimensional and multi-granular social simulation context. Regarding agent initialization, a human-like agent with cognitive self-evolution capabilities is constructed for each individual. The individual model is parameterized as a cognitive entity composed of a memory graph structure (immediate memory, short-term memory, and long-term memory). Its initial state is achieved through the injection of a small amount of prior knowledge and role characteristics, while also incorporating a multimodal perception coding module and a personality characteristic constraint mechanism.

[0026] 2. Dynamic Memory Graph-Driven Contextual Perception and Cognitive Construction. To enhance the agent's adaptability to dynamic task environments, a dedicated memory controller is introduced to coordinate knowledge at different levels in the memory pool and dynamically embed it into a multimodal large model. When perceiving external environmental input, this invention abandons the traditional context splicing method and instead extends long-term and short-term memory cues into the individual's prior cognitive space. In the specific memory processing process, the memory controller intercepts external input stimuli, generates transient memory cues to replace the inherent input sequence of the large model, and combines them with fixed short-term and long-term memory cues. This cue optimization mechanism enables the multimodal large model to accurately focus on and acquire transient perceptual knowledge in the current task under the continuous guidance of retaining the individual's historical memory knowledge (long-term and short-term memory cues), achieving deep fusion and joint reasoning of knowledge at different levels at the model input end.

[0027] Specifically, this invention abandons the traditional method of concatenating contextual text, and the system uses long-term memory prompts. (Such as ingrained beliefs, long-term experiences) and short-term memory cues from the previous moment. As a priori cognitive vector, it is embedded into the cognitive representation space of the individual's multimodal large model. When perceiving multimodal external input... At this time, there is no need to perform traditional memory similarity retrieval; the current input is directly mapped to a high-dimensional instantaneous memory cue. And serve as dynamic cognitive input for the current moment: ; in, For the perceptual coding network, a pre-trained encoder with frozen parameters is used to extract external input. The multimodal representations are then mapped to a continuous cue space that is strictly aligned with the input dimension of the large model through a lightweight, learnable projection layer.

[0028] At the feature layer, the newly generated transient memory cue and the fixed prior cue are concatenated at the vector level and aligned in dimensions to construct a dynamic joint cue space. : .

[0029] in, This represents the dynamic memory graph fusion function, used to construct a dynamic joint cue space. (Multimodal large model) Under the continuous guidance of preserving the underlying constraints of individual historical memory knowledge (long-term memory cues and short-term memory cues), it can accurately focus on the current task context using self-attention mechanisms, thereby directly generating cognitive responses that conform to psychological consistency and contextual sensitivity. : .

[0030] 3. Efficient modeling of transient memory based on Bayesian distribution alignment.

[0031] Individual memory is not a static mapping result, but a dynamic process that continuously evolves with social interactions. By introducing cognitive feedback loops and adaptive evolutionary mechanisms, a continuous cognitive update and preference transfer path for individuals in complex social contexts is constructed. Immediate memory needs to simultaneously possess high stability in preserving key historical features and strong plasticity in rapidly adapting to new tasks. To overcome this "stability-plasticity" dilemma, the memory controller introduces a Bayesian distribution alignment mechanism when learning and updating immediate memory. Specifically, this step constructs a Bayesian neural network to simulate the perceptual distribution of immediate memory under different dynamic tasks or evolutionary stages through adversarial training. When the agent faces new task stimuli, the system optimizes the posterior distribution of the immediate memory cue to align it with the distribution characteristics of the target task in the parameter space. This mechanism ensures that the immediate cue can efficiently model and store the immediate experience of the current task without compromising the stability of existing cognition due to drastic parameter oscillations, thus achieving efficient representation of immediate knowledge. The tasks in this invention may include: understanding and judging news events, forming opinions or attitudes about social media content, responding to comments or behaviors of other intelligent agents, choosing whether to forward or spread information, judging the credibility of information, adjusting social relationships, and making behavioral decisions in different event scenarios.

[0032] Specifically, in the efficient modeling of transient memory based on Bayesian distribution alignment, this invention introduces transient memory cues into the probabilistic model, constructing a distribution alignment loop based on variational Bayesian and adversarial training. Simultaneously, it utilizes a Bayesian neural network to integrate the transient memory cues... Modeled as a Gaussian distribution: ; The central representation for stable short-term memory represents the memory form that the system considers most reasonable and stable. This represents the uncertainty of the current instantaneous memory, used to measure the magnitude of the system's uncertainty about this memory. , ; and There are two projection layers. The former outputs stable memory centers, and the latter outputs memory uncertainty.

[0033] Stable transient memory representations are obtained through sampling: .

[0034] It should be noted that the object modeled by this Gaussian distribution is not the instantaneous memory cue of the current input. It is not itself, but a stable transient memory representation verified by cognitive response. The system uses the current instantaneous memory prompts. and cognitive response Generate stable transient memories posterior distribution Its mean is The variance is In other words, the system does not directly treat the current perceived result as memory content to be stored, but rather determines, under the joint constraints of the current input and cognitive response, what should ultimately be written into the memory pool as stable memory. Where are they most likely to be distributed, thus seeking the optimal balance between perceived uncertainty and cognitive consistency.

[0035] When updating the posterior distribution of transient memories, the system constructs a ternary joint loss function that includes task adaptation, distribution stability, and adversarial alignment. : ; in, For task bias loss; KL divergence term Strictly constrain the current approximate posterior distribution With historical prior distribution The parameter offset is used to prevent catastrophic amnesia caused by severe oscillations.

[0036] To counteract training loss, a discriminator network is introduced. Distinguishing between instantaneous memory sampling features and real-world environmental features of the target task using a game-theoretic approach: .

[0037] After optimization using the joint loss function, a posterior representation of instantaneous memory that is consistent with the current task distribution and maintains the continuity of historical cognition is obtained: ; in, This represents a stable transient memory representation after Bayesian distribution alignment. It retains the consistency constraints between the information features of the current external input and the historical cognitive state, and serves as the input for subsequent cross-task knowledge mining modules.

[0038] 4. Cross-task knowledge mining and multi-level cognitive integration.

[0039] The self-evolution of individual cognition relies not only on the acquisition of instantaneous experiences, but more importantly, on the sedimentation of short-term experiences into long-term knowledge. During the instantaneous memory learning process for a new task, historical instantaneous memories already stored in the memory pool are activated, and this knowledge is used to assist the current model in recalling and reconstructing historical experiences. Subsequently, this invention initiates a cross-task knowledge mining module to perform semantic refinement and structured extraction on instantaneous memories that have been aligned by Bayesian distribution and verified as high-value through multiple rounds of interaction. Through a cross-task mapping mechanism, these verified instantaneous experience crystallizations are dynamically updated and integrated into short-term memory cues and long-term memory cues. This step achieves top-down integration of knowledge from the instantaneous state to the long-term state, enabling the agent to break the isolation of a single task and truly realize the continuous accumulation and transfer of knowledge.

[0040] Specifically, in the cross-task knowledge mining and multi-layered cognitive integration module, this invention constructs a hierarchical cognitive evolution mechanism composed of instantaneous memory, short-term memory, and long-term memory. The continuous evolution of individual cognition depends not only on instantaneous learning under the current external input, but also on the accumulation, generalization, and transfer of historical experience in different task scenarios.

[0041] After the agent completes the current task interaction, it will use the stable transient memory prompts aligned with the Bayesian distribution. and the corresponding task scenarios Environmental conditions and cognitive response They are collectively encapsulated as historical experience units. Stored in the history memory pool middle: .

[0042] By recording relevant information such as the task scenario, environmental state, and cognitive response, a cross-task retrieval mechanism can be used to activate a set of historical instantaneous memories that are potentially related to the current task during subsequent task execution. .

[0043] in, This represents the set of candidate experiences retrieved from the historical memory pool. The retrieval function is driven by a Large Model (LLM). Semantic induction and knowledge compression are performed on candidate experiences to extract stable cognitive patterns shared across tasks. ; in, This represents a knowledge mining function used to identify high-level semantic features such as recurring behavioral patterns, value preferences, decision-making criteria, and cognitive tendencies across different tasks. This indicates that knowledge is shared across tasks.

[0044] Subsequently, a knowledge score is constructed based on the importance of the knowledge, its frequency of recurrence, and its contribution to the task: ; in, This indicates the frequency of knowledge occurrence in historical tasks. This indicates the degree to which knowledge contributes to the success of the task. This indicates the consistency between knowledge and an individual's existing cognitive system. , , These are the cognitive weight parameters.

[0045] The calculation method is as follows: ; in, Indicates the first An intelligent agent at a historical moment Stored in the historical memory pool Historical experience unit, express Corresponding cross-task knowledge sharing, Represents the similarity function. Indicates the time decay weight; This represents the response semantic encoding function.

[0046] The calculation method is as follows: ; in To retrieve the relevant subset of historical experiences that have been hit, The preset similarity threshold; This represents the reward value for historical missions. This represents the normalization function.

[0047] The calculation method is as follows: ; in, Represents cosine similarity. This serves as a short-term memory cue for the previous moment.

[0048] When the knowledge score exceeds a set threshold, a cognitive integration mechanism is triggered, synchronously updating high-value knowledge to both short-term and long-term memory cues. , ; in, and These represent the short-term memory update function and the long-term memory update function, respectively. The former is used to reinforce knowledge related to recent tasks, while the latter is used to solidify long-term stable cognitive structures.

[0049] 5. Environmental updates and cognitive self-evolution and iteration.

[0050] After completing the aforementioned cross-layer knowledge integration, the agent outputs its current cognitive judgment, stance decision, or interactive behavior based on the updated multimodal large model and applies it to the multi-agent simulation environment. The environment server captures the behavioral outputs of all individuals, dynamically updates the global situation and event evolution state, and generates a new round of heterogeneous information stimuli to feed back to each agent. With the continuous input of environmental information, the agent re-triggers the flow process from steps S3 to S6, relying on the continuous updates of parameterized memory and the continuous guidance of prompt optimization to form a cognitive self-evolutionary closed loop of "environmental stimulus - instantaneous perception - cross-layer integration - decision output". This two-way drive of environmental evolution and individual cognitive evolution provides highly anthropomorphic and continuously explanatory technical support for the long-term deduction and cognitive dynamic evolution of complex systems.

[0051] This invention constructs a unified dynamic cognitive representation space, maps external inputs to transient memory cues through a perceptual coding network, and performs vector-level fusion and dimensional alignment of the memory graph at the feature layer to form a joint cue space, which drives a multimodal large model to generate cognitive responses. It achieves end-to-end context modeling based on the unified representation space, enabling the model to complete semantic aggregation and context perception within a structured cue space. A method and system for agent cognitive self-evolution based on dynamic memory graphs is also presented.

[0052] This invention introduces transient memory cues into a probabilistic modeling framework, using a Bayesian neural network to model the transient memory representation in a distributed manner, ensuring it conforms to a Gaussian distribution. The system achieves end-to-end optimization through a reparameterization method and constructs a joint optimization objective function by combining task adaptation loss, KL divergence constraints, and adversarial training loss, thereby achieving distribution alignment and stabilization of transient memories.

[0053] This invention constructs a cross-task knowledge evolution mechanism based on a historical memory pool. Distributed and aligned instantaneous memory representations are stored in the historical memory pool, and a cross-task retrieval function extracts a subset of memories relevant to the current context from historical experience. The system further uses a knowledge mining function to perform semantic compression and structured abstraction on the retrieval results, forming stable cross-task cognitive knowledge.

[0054] Example: In each round of simulated event interaction, the agent first receives multimodal external input information, performs unified access and format standardization processing on the input data, and maps text, images, or structured behavioral data into a unified feature representation. Simultaneously, the agent reads long-term and short-term memory from its internal memory modules. Long-term memory stores the agent's stable preferences, belief structures, and historical experiences formed over a long period; short-term memory stores the agent's temporary cognitive states and local behavioral tendencies formed in recent rounds of interaction. Finally, the agent unifies the current external input with its long-term and short-term memory to form a complete set of cognitive inputs for the current moment.

[0055] To ensure the agent can accurately understand and respond to multimodal information, semantic encoding is performed on the current multimodal external input. High-dimensional semantic features are extracted using a pre-trained multimodal coding network to generate a transient memory representation, which characterizes the current environmental state and task context. Subsequently, the transient memory representation is fused and aligned with long-term and short-term memory, constructing a joint cognitive representation space through a feature-level fusion mechanism. In this space, memory information from different time scales interacts collaboratively, forming a unified cognitive representation that includes historical constraints and the current context. This joint cognitive representation serves as the input foundation for subsequent reasoning modules, supporting individual cognitive judgments and attitude generation.

[0056] Based on a joint cognitive representation space, the agent generates its current cognitive response through cognitive reasoning. Simultaneously, to enhance the stability and robustness of the instantaneous cognitive representation, a Bayesian distribution modeling mechanism is introduced, modeling the instantaneous memory representation as a probability distribution and parameterizing it through variational inference. The system constrains the instantaneous cognitive distribution by jointly optimizing the task objective function, historical distribution constraints, and adversarial consistency constraints, enabling the agent to maintain consistency with its historical cognitive structure while adapting to changes in current input, thus obtaining a stable instantaneous cognitive representation. This stable representation not only preserves current environmental information but also suppresses cognitive shifts caused by noise or short-term disturbances, serving as the input basis for subsequent memory updates and knowledge storage.

[0057] The stabilized instantaneous cognitive representations are written into the historical memory pools of each agent's memory module, and their corresponding task scenarios, environmental states, and cognitive output information are recorded simultaneously, thus forming structured cognitive experience data. During subsequent task execution, the agent retrieves relevant experience sets from the historical memory pools based on the current input context, and performs semantic induction and structured abstraction on the retrieval results to extract stable cognitive patterns shared across tasks, forming high-level knowledge representations. Simultaneously, the importance of the extracted knowledge is assessed based on its frequency of occurrence in historical tasks, its contribution to task completion, and its consistency with existing cognitive structures, and high-value knowledge is selected for subsequent updates.

[0058] When the agent determines that certain knowledge meets the update conditions, it writes high-value knowledge into its short-term memory and long-term memory modules respectively. The agent uses short-term memory to update recent behavioral strategies, enabling it to quickly adapt to changes in the current environment and rapidly adjust local behavioral patterns. Simultaneously, the agent uses long-term memory to solidify stable cognitive structures, gradually forming consistent long-term individual cognitive preferences and achieving the gradual evolution of these preferences. In the next round of interaction, the agent reloads the updated long-term and short-term memories and participates in cognitive computation along with new external inputs. By reloading the updated memory map for computation, a cyclical evolutionary process of "input-cognitive construction-stabilization-storage-retrieval and update-reinput" is formed. Through this mechanism, the individual agent can achieve a gradual evolution of its cognitive state during continuous interaction, enabling it to possess cross-task knowledge transfer capabilities and long-term consistent cognitive modeling capabilities.

[0059] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0060] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0061] This invention also provides a system. The system can be a system that uses software (applications), modules, components, servers, clients, etc., employing the methods described in the embodiments of this specification, combined with necessary hardware implementation. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can be found in the implementations of the foregoing methods, and repeated details will not be elaborated upon. Although the system is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0064] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for the cognitive self-evolution of intelligent agents based on dynamic memory maps, characterized in that, Includes the following steps: S1, construct a memory map for each agent that includes long-term memory, short-term memory, and transient memory; S2, select long-term memory nodes and short-term memory nodes from the memory map, and generate long-term memory cues and short-term memory cues respectively through aggregation, encoding and cue mapping; S3 maps external input to instantaneous memory prompts; Immediate memory cues are concatenated with long-term memory cues and short-term memory cues at the feature layer and aligned with their dimensions to form a joint cue space, which drives the multimodal large model to generate cognitive responses; S4, model the transient memory cue as a probability distribution, and use the Bayesian distribution alignment mechanism to perform parameter constraints and stabilization to obtain a stable transient memory representation; S5 encapsulates stable transient memory representations into historical experience units and stores them in the historical memory pool. Through cross-task retrieval and knowledge mining, cross-task shared knowledge is extracted, and this shared knowledge is used to update short-term memory cues and long-term memory cues. S6, the cognitive response is output to the multi-agent simulation environment, the environment server captures the behavioral output of each agent, dynamically updates the global situation and generates a new round of external input feedback to each agent, and returns to step S3.

2. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 1, characterized in that, The process of mapping external input to instantaneous memory prompts specifically includes: Through perceptual coding network Extracting external inputs from multimodal systems The high-dimensional representation is then mapped through a learnable projection layer to a continuous cue space aligned with the input dimension of the multimodal large model, generating transient memory cue. ,in, Indicates the agent index. This represents the index at time t.

3. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 1, characterized in that, The step of concatenating and aligning the immediate memory cues with long-term memory cues and short-term memory cues at the feature layer to form a joint cue space specifically includes: Through dynamic memory map fusion function Constructing a federated suggestion space ,in, For long-term memory cues, This serves as a short-term memory cue for the previous moment. This serves as a prompt for the current immediate memory.

4. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 1, characterized in that, The process of modeling transient memory cues as a probability distribution and using a Bayesian distribution alignment mechanism for parameter constraints and stabilization to obtain a stable transient memory representation specifically includes: Current instantaneous memory prompt Modeled as an approximate posterior distribution : ； For the stable representation of immediate memory, The cognitive response generated by the agent based on the current input. It follows a Gaussian distribution. To stabilize the central representation of immediate memory, The uncertainty of current instantaneous memory; Sampling yields a stable instantaneous memory representation : ； in, , For random perturbations sampled from the standard normal distribution, It is the identity matrix. This is element-wise multiplication; Constructing a ternary joint loss function Update the posterior distribution of the immediate memory: ； in, For task deviation loss, To constrain the approximate posterior distribution With historical prior distribution KL divergence term of parameter offset, To counteract training losses, , These are the weighting coefficients. This serves as a reminder for immediate memory from the previous moment; After optimization using the joint loss function, a stable instantaneous memory representation aligned with the Bayesian distribution is obtained. .

5. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 4, characterized in that, The adversarial training loss specifically includes: Introducing a discriminator network It distinguishes between instantaneous memory sampling features and target task environmental features through game theory, and combats training loss. for: ； in, The distribution of real-world environmental features for the target task. Expressing expectations, Represents external input samples It conforms to the distribution of real-world environmental characteristics of the target task.

6. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 4, characterized in that, Task deviation loss The calculation formula is: ； in, Indicates the target response for the current task. This represents the response semantic encoding function. This represents the L2 norm.

7. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 1, characterized in that, The process of encapsulating stable transient memory representations into historical experience units and storing them in a historical memory pool, followed by cross-task retrieval and knowledge mining to extract cross-task shared knowledge, specifically includes: Representing stable transient memory and the corresponding task scenarios Environmental conditions and cognitive response They are collectively encapsulated as historical experience units. : ； Units of historical experience Stored in the historical memory pool ,Right now ; By retrieval function From the pool of historical memory Activate the set of candidate experiences that are potentially relevant to the current task. : ； Through knowledge mining functions Semantic induction and structural abstraction are performed on the candidate experience set to extract cross-task shared knowledge. .

8. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 1, characterized in that, The method of updating short-term and long-term memory cues using this shared knowledge specifically includes: Based on cross-task knowledge sharing Frequency of occurrence in historical missions The degree of contribution to the completion of the task and consistency with existing cognitive systems Construct knowledge scoring When the knowledge score exceeds a preset threshold, the cross-task shared knowledge will be... Synchronized updates to short-term memory tips and long-term memory tips : ， ； in, These are the short-term memory cues from the previous moment and the current short-term memory cues, respectively. For short-time memory update functions, This is a function for updating long-term memory.

9. The method for agent cognitive self-evolution based on dynamic memory graphs according to claim 8, characterized in that, Knowledge score The calculation method is as follows: ； in, , , These are weight parameters; The calculation method is as follows: ； in, Indicates the first An intelligent agent at a historical moment Stored in the historical memory pool Historical experience unit, express Corresponding cross-task knowledge sharing, Represents the similarity function. Indicates the time decay weight; This represents the response semantic encoding function; This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and takes the value 0 if the condition within the parentheses is not met. The calculation method is as follows: ； in To retrieve the relevant subset of historical experiences that have been hit, The preset similarity threshold; This represents the reward value for historical missions. Represents the normalization function; The calculation method is as follows: ； in, Represents cosine similarity. This serves as a short-term memory cue for the previous moment.

10. A self-evolving cognitive system for intelligent agents based on dynamic memory maps, used to implement the method described in any one of claims 1 to 9, characterized in that, include: The memory map construction module is used to construct a memory map for each agent, which includes long-term memory, short-term memory, and transient memory. The memory cue generation module is used to filter long-term memory nodes and short-term memory nodes from the memory graph, and generate long-term memory cues and short-term memory cues respectively through aggregation, encoding and cue mapping; The joint cognition construction module is used to concatenate and align the vectors of immediate memory cues, long-term memory cues, and short-term memory cues at the feature layer to form a joint cue space, which drives the multimodal large model to generate cognitive responses. The Bayesian distribution alignment module is used to model transient memory cues as a probability distribution, and obtains a stable transient memory representation by parameterizing and stabilizing them through the Bayesian distribution alignment mechanism. The cross-task knowledge integration module is used to encapsulate stable transient memory representations into historical experience units and store them in the historical memory pool. Cross-task shared knowledge is extracted through cross-task retrieval and knowledge mining, and the shared knowledge is used to update short-term memory cues and long-term memory cues. The environment interaction and iterative cycle module is used to output the cognitive response to the multi-agent simulation environment. The environment server captures the behavioral output of each agent, dynamically updates the global situation, generates a new round of external input feedback to each agent, and triggers the joint cognitive construction module to be executed again.

Images