Social subject memory simulation system and method based on large language model

Abstract

The invention relates to a social subject memory simulation system and method based on a large language model, and belongs to the technical field of computers. According to the method, online and offline multi-source information fusion is realized, and behavior and situation characteristics of social subjects can be comprehensively described; a large language model semantic comprehension capability is introduced, so that the intelligent level of memory retrieval and information fusion is improved; dynamic dump and index retrieval of long-term memory are supported, so that subjects can keep semantic coherence and behavior consistency in multiple rounds of interaction; the authenticity of information spreading and interaction between social subjects is enhanced, and technical support is provided for an intelligent social system, virtual human interaction and social simulation.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a social subject memory simulation system and method based on a large language model. Background Technology

[0002] With the widespread adoption of social networks, user behavior in the virtual and real worlds is increasingly intertwined. The effective dissemination and integration of online and offline information has become a crucial research direction for understanding user behavior, optimizing social experiences, and enhancing system intelligence. Online information typically originates from users' online activities, social interactions, content browsing, and posting behaviors; while offline information includes users' daily lives, environmental perceptions, emotional states, and actual activities. How to organically integrate these two types of heterogeneous information and construct a coherent, long-term subjective memory system is one of the key issues in current research on the intelligence of social network technology.

[0003] Existing technologies largely focus on the storage and analysis of online behavioral data, such as capturing user preferences through log analysis, recommendation algorithms, or user profiling. However, these methods typically overlook the dynamic changes in offline information and the impact of social context on behavioral decisions, resulting in a lack of contextual continuity when generating personalized content or engaging in multi-turn interactions. Especially in the context of the widespread application of large language models in social scenarios, traditional single-source information modeling methods struggle to support the long-term memory and cross-scenario reasoning needs of social participants. The fusion of online and offline information can not only more comprehensively present user behavior patterns but also provide strong support for tasks such as personalized recommendations and emotional support.

[0004] Social agents are the core modules simulating user behavior on social networks, possessing both online and offline attributes, including but not limited to social accounts, interests, behavioral habits, and location awareness. Each social agent simulates a user's browsing, interaction, and posting behaviors to showcase personalized social dynamics, supported by a memory system for multi-turn dialogues and long-term interactions. Furthermore, traditional information dissemination and memory management mechanisms are often fragmented, lacking a unified storage architecture and interaction interface for online and offline information. This results in the system lacking a complete understanding of the context when generating personalized content, leading to insufficient relevance between the generated output and the user's actual needs and interests. Summary of the Invention

[0005] (a) Technical problems to be solved The technical problem to be solved by this invention is to provide a method for simulating the online and offline memory of social subjects, which is insufficient in the existing technology due to the lack of online and offline information integration and dynamic memory mechanism. This method enables the dynamic collection, integration, dissemination and storage of multi-source information of social subjects, thereby supporting the continuity of multi-turn dialogues and the simulation of long-term social behavior.

[0006] (II) Technical Solution To address the aforementioned technical problems, this invention provides a social agent memory simulation system based on a large language model, comprising: The information collection and storage module is used to collect online activity information and offline environmental information of social subjects, and store them in a structured way to form long-term memory; The memory fusion module is used to integrate the online activity information with the offline environmental information to achieve multi-source information fusion and form a comprehensive memory of the social subject; The information dissemination and memory transfer module is used to initiate and maintain multi-round dialogues among multiple social subjects based on comprehensive memory to achieve information dissemination, and dynamically transfer the interactive information during the dissemination process to the comprehensive memory to achieve dynamic updating and synchronization of long-term memory.

[0007] (III) Beneficial Effects This invention provides a social subject memory simulation system and method based on a large language model, which enables the exchange and diffusion of information between subjects through dialogue, and dynamically summarizes, stores, retrieves, and integrates memory content. This method effectively simulates the information dissemination and long-term memory accumulation process in complex social scenarios, providing a novel technical solution for the coherence of multi-turn dialogues, the naturalness of subject behavior, and the intelligent development of social networks.

[0008] Compared with existing technologies, this invention achieves the fusion of online and offline multi-source information, which can comprehensively depict the behavior and contextual characteristics of social subjects; it also introduces the semantic understanding capabilities of a large language model, which improves the intelligence level of memory retrieval and information fusion; it supports the dynamic dumping and index retrieval of long-term memory, enabling subjects to maintain semantic coherence and behavioral consistency in multiple rounds of interaction; and it enhances the authenticity of information dissemination and interaction between social subjects, providing technical support for intelligent social systems, virtual human interaction, and social simulation. Attached Figure Description

[0009] Figure 1 This is a schematic diagram illustrating the overall implementation principle of the method of the present invention. Detailed Implementation

[0010] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0011] This invention relates to a social agent memory simulation system and method based on a large language model. Specifically, this invention utilizes artificial intelligence and natural language processing technologies to simulate agent behavior, memory formation, and information dissemination processes in virtual social networks, thereby achieving intelligent modeling and analysis of social interactions and group dynamics. The main technical challenge of this invention lies in achieving seamless integration and dissemination of online and offline information, and storing key actions and decisions of the agent in long-term memory through a memory dumping mechanism, enabling subsequent dialogues or decisions to be generated based on a complete memory context. Simultaneously, an effective information fusion mechanism needs to be established to combine online and offline information to form a unified behavioral trajectory, thereby supporting the coherence and authenticity of the agent in multiple scenarios. This innovative memory dumping and fusion technology lays an important foundation for the intelligence and naturalization of social network agents.

[0012] This invention improves the intelligence and personalization of the system by constructing a social subject information collection module, a memory retrieval and fusion module, and an information dissemination and memory transfer module, thereby achieving cognitive consistency and behavioral coherence of social subjects in complex social scenarios.

[0013] (I) Overall System Structure According to embodiments of the present invention, a social agent memory simulation system based on a large language model is provided. For example... Figure 1 As shown, the system mainly includes: The information collection and storage module is used to collect online activity information and offline environmental information of social subjects, and store them in a structured way to form long-term memory; The memory fusion module is used to integrate the online activity information with the offline environmental information to achieve multi-source information fusion and form a comprehensive memory of the social subject; The information dissemination and memory transfer module is used to initiate and maintain multi-round dialogues among multiple social subjects based on comprehensive memory to achieve information dissemination, and dynamically transfer the interactive information during the dissemination process to the comprehensive memory to achieve dynamic updating and synchronization of long-term memory, simulating the accumulation of long-term memory and the influence of behavior.

[0014] The system uses a large language model to simulate multi-turn dialogues and manage long-term memory of social subjects, thereby improving the coherence and realism of subject interactions.

[0015] (II) Information Collection and Storage Module In the description of this invention, the information collection and storage module is used to collect behavioral data of social subjects from both online and offline dimensions, and to perform unified formatting and structured storage to support subsequent semantic retrieval and memory updates, including an online information collection submodule and an offline information collection submodule.

[0016] (1) Online information collection submodule The online information collection submodule is used to acquire online behavioral data of social media users from social media platforms and store this information in a structured manner for social media user behavior modeling and information dissemination. The specific implementation steps of this online information collection submodule are as follows: S11: Online activity collection; real-time acquisition of online activity data from different dimensions through social platform interfaces, including browsing history, interaction history, and published content of social entities; S12: Formatted storage of behavioral data; preprocess the collected behavioral data to extract key fields such as timestamp, source, topic tags, and interaction behavior; then generate unique identifiers for fast retrieval and duplicate detection; finally, store the structured results in an online behavioral memory database in a unified format; S13: Data deduplication and update mechanism; deduplicating and updating data entries in the online behavior memory database, comparing the unique identifiers of newly collected data with those of existing data, and deleting duplicate records; for existing entries, updating their interaction status or additional information based on the latest timestamp.

[0017] (2) Offline information collection submodule The offline information collection submodule simulates the daily behavior and activity environment of social subjects in real-world scenarios, thereby generating offline behavior records and activity plans, achieving dynamic modeling of real-world behavior. Specifically, this offline information collection submodule implements the following steps: S14: Simulate offline behavior; generate reasonable daily activity plans based on the personality traits, interests, preferences, and time arrangements of the social subjects, including work, leisure, social, or travel activities; further simulate offline interactions between subjects, such as gatherings, meetings, or community participation events; output behavioral trajectories and event sequences that conform to the characteristics of the subjects through the action rules module; S15: Generate activity records; store the simulated activity content (including time, location, participants, behavioral descriptions, etc.) in an offline behavioral memory database as an important part of the social subject's long-term memory, providing data support for subsequent information fusion and memory retrieval.

[0018] (III) Memory Fusion Module In the description of this invention, the memory fusion module is used to realize the dynamic retrieval of memory information and the fusion of online and offline data during multi-turn dialogues, ensuring the semantic consistency and behavioral coherence of the subject, including a memory retrieval submodule and an information fusion submodule.

[0019] (1) Memory retrieval submodule The memory retrieval submodule dynamically retrieves online and offline behavioral memory databases (online and offline long-term memories) of social participants to support the generation of current dialogues or behaviors. Specifically, it extracts memory information related to the topic or scenario of the current dialogue from online and offline behavioral memory databases (long-term memory databases) through semantic matching and priority algorithms, and naturally integrates the retrieved memory information into the dialogue to enhance its coherence and contextuality. Its execution steps are as follows: S21: Determine retrieval needs; generate memory retrieval requests based on the topic of the conversation, context, or behavioral needs; for example, when the conversation involves "recent activities" or "emotional state," automatically determine the corresponding memory category; S22: Perform memory retrieval; retrieve relevant content from online and offline behavioral memory databases based on semantic matching algorithms; combine keyword matching, semantic embedding vector similarity, and memory priority (time, relevance, importance) to filter and sort, and extract the most valuable memory information; S23: Generate search results; convert the retrieved memory information into natural language expressions that can be embedded in dialogue, so that the generated memory information is naturally connected with the current context; at the same time, record the results of this search, and cite the context and user feedback, so as to optimize the search strategy in the future.

[0020] (2) Information Fusion Submodule The information fusion submodule integrates the retrieval results (online and offline information of social subjects) from the memory retrieval submodule to generate a unified behavioral context, providing support for subsequent content generation and memory dumping. Its execution steps are as follows: S24: Integrate online and offline information, sort the retrieval results of the memory retrieval submodule based on the timeline, and use the semantic disambiguation capability of the large language model to eliminate duplication and conflict; further establish cross-source semantic associations for the sorted information, such as linking online comment behavior with offline participation in activities to form a unified event chain.

[0021] S25: Generate behavioral context; extract key elements from the event chain (including events, emotions, tasks, and participants) and construct a multi-level semantic context to support dialogue generation and memory updates.

[0022] (iv) Information dissemination and memory transfer module The information dissemination and memory dumping module facilitates the natural transmission of information between social participants and performs semantic summarization and long-term storage of important content after a conversation ends to support contextual continuity in subsequent interactions. This module comprises an information dissemination submodule and a memory dumping submodule.

[0023] (1) Information dissemination submodule The information dissemination submodule combines information from online and offline behavioral memory databases to generate natural dialogue content that closely resembles real-life contexts, promoting interaction and information diffusion among social participants. The specific steps are as follows: S31: Information filtering; based on the dialogue context, filter the most valuable information from online and offline behavioral memory databases; use a large language model to evaluate the relevance, interest and context matching of the information, and prioritize content that is of interest to both parties in the dialogue or has discussion value.

[0024] S32: Dialogue content generation; The large language model combines the language style and social preferences of the social subjects to personalize the filtering results of the information to be disseminated, that is, to generate language responses that are logically consistent and semantically natural based on the filtering results and the current topic, and then to verify the consistency and grammar of the generated language responses. S33: Information dissemination execution; insert the verified language response content into the multi-level semantic context generated by the memory retrieval and fusion module, initiate and maintain multi-round dialogues, realize information diffusion and behavioral interaction between social subjects; and record information such as disseminated content, time, and recipients to support subsequent memory updates.

[0025] (2) Memory dump submodule The memory dump submodule is used to summarize, refine, and store the interaction content in online and offline behavioral memory databases after a dialogue or event ends, using a large language model. This enables dynamic memory accumulation and sharing among social participants. Specifically, it includes: S34: Generate a dialogue summary; after each round of dialogue, call the large language model to perform semantic analysis, extract key content (such as activity arrangements, interaction results, emotional expressions, etc.), and generate structured summary content, with timestamps and identifiers of participating social entities attached; S35: Store summary content; store summary content in the online and offline behavioral memory database of the participants in the social interaction, and dynamically adjust memory priority based on the importance, time and frequency of citation; S36: Memory synchronization and optimization; in subsequent interactions, key memories are gradually spread to the long-term memories of other social subjects to achieve information sharing and the construction of a social memory network; at the same time, redundant or low-relevance memory data is cleaned up regularly to maintain the efficient operation of the system.

[0026] The method of the present invention based on the above system includes the following main steps: (1) Construct a module for collecting and storing online and offline information of social subjects; (2) Construct a memory retrieval module and an information fusion module to realize the integration and sharing of online and offline memories of social subjects; (3) Construct an information dissemination module and a dynamic dumping mechanism for memory content to support the continuity of multi-round dialogues and long-term interaction simulation of social subjects.

[0027] The specific process of the method of the present invention based on the above system is as follows: Step 1. Online and offline information collection and storage by social entities This step involves collecting behavioral data and environmental information from various sources and establishing a unified, structured storage system. Specifically: Step 1.1) Online information collection by social media participants ① Real-time acquisition of online activity data from different users through social platform interfaces, including browsing history, interaction history, and published content; ② Establish a unified data format specification and storage structure based on the attributes and types of the collected online information, and store it in a memory database to facilitate efficient retrieval and fusion processing in the future.

[0028] Step 1.2) Offline information collection of social subjects ① Collect daily activities and environmental data of social subjects through the perception module, including activity arrangements, location awareness, emotional state and dynamic events; ② The collected offline information is dynamically recorded in a structured or semantic form into the social subject's daily plan and status database for subsequent behavioral decisions, context generation, and information integration.

[0029] Step 2. Social Subject Memory Retrieval and Information Fusion Step 2.1) Social Subject Memory Retrieval ① Based on the current dialogue topic and the needs of the interaction task or scenario, dynamically retrieve related memory fragments from the social subject memory database; ② By combining historical memory to generate a coherent contextual dialogue flow, the continuity of language and behavior of real users in long-term social interactions is simulated.

[0030] Step 2.2) Integration of online and offline information for social entities ① Utilize memory retrieval mechanisms to obtain online information (social behavior, interactive content, etc.) and offline information (activity status, emotional expression, etc.) related to the current scene or topic; ② Based on the contextual understanding and semantic modeling capabilities of large language models, multi-source information is semantically fused to form a complete and comprehensive context for dialogue behavior decision-making, providing support for subsequent information dissemination and content generation.

[0031] Step 3. Information dissemination and memory transfer of social subjects: Step 3.1) Information dissemination by social subjects ① The natural dissemination of information is achieved through dialogue between social subjects. The large language model selects the most valuable content for dissemination and embeds it into the dialogue context based on the dialogue context, semantic association and social motivation. ② The system dynamically integrates online and offline information to generate dialogue content that closely resembles real-world scenarios, thereby ensuring the authenticity of information dissemination and the consistency of subject behavior interaction.

[0032] Step 3.2) Social Subject Memory Dump ① Summarize and identify key information in the dialogue, such as planning, behavioral decisions, and emotional expressions, and transfer it to the subject's long-term memory; ② Create time and topic indexes for the memorized content to facilitate subsequent retrieval and recall.

[0033] As can be seen, this invention proposes a social agent memory simulation method based on a large language model. In a multi-agent social system, this method combines agents' online information (such as social media activities and web browsing history) with offline information (such as daily plans and environmental perception data), disseminating information between agents through dialogue and ultimately influencing their behavioral decisions. Through innovative information dissemination and memory dumping mechanisms, this invention simulates the process of information acquisition, information sharing, and long-term memory accumulation in human social networks, and has broad application prospects.

[0034] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A social agent memory simulation system based on a large language model, characterized in that, include: The information collection and storage module is used to collect online activity information and offline environmental information of social subjects, and store them in a structured way to form long-term memory; The memory fusion module is used to integrate the online activity information with the offline environmental information to achieve multi-source information fusion and form a comprehensive memory of the social subject; The information dissemination and memory transfer module is used to initiate and maintain multi-round dialogues among multiple social subjects based on comprehensive memory to achieve information dissemination, and dynamically transfer the interactive information during the dissemination process to the comprehensive memory to achieve dynamic updating and synchronization of long-term memory.

2. The system as described in claim 1, characterized in that, The information collection and storage module is used to collect behavioral data of social subjects from both online and offline dimensions, and to perform unified formatting and structured storage to support subsequent semantic retrieval and memory updates.

3. The system as described in claim 1, characterized in that, The information collection and storage module includes an online information collection submodule and an offline information collection submodule; (1) Online information collection submodule The online information collection submodule is used to acquire online behavior data of social media users from social media platforms, and to store this data in a structured manner for social media user behavior modeling and information dissemination. The specific implementation steps are as follows: S11: Online activity collection; real-time acquisition of online activity data from different dimensions through social platform interfaces, including browsing history, interaction history, and published content of social entities; S12: Formatted storage of behavioral data; The collected behavioral data is preprocessed to extract key fields such as timestamp, source, topic tags, and interaction behavior; A unique identifier is then generated for rapid retrieval and duplicate detection; finally, the structured results are stored in an online behavioral memory database in a unified format. S13: Data deduplication and update mechanism; deduplicates and updates data entries in the online behavior memory database, compares the unique identifiers of newly collected data with those of existing data, and deletes duplicate records; for existing entries, updates the interaction status based on the latest timestamp; (2) Offline information collection submodule The offline information collection submodule is used to simulate the daily behavior and activity environment of social subjects in real-world scenarios, thereby generating offline behavior records and activity plans of the subjects, realizing dynamic modeling of real-world behavior. The specific implementation steps are as follows: S14: Simulate offline behavior; generate daily activity plans based on the personality traits, interests, preferences, and schedules of the social subjects, including work, leisure, social, or travel activities; further simulate offline interactions between subjects, including gatherings, meetings, or community participation events; output behavioral trajectories and event sequences that conform to the characteristics of the subjects; S15: Generate activity record; The simulated offline activities are stored in an offline behavior memory database as part of the social subject's long-term memory.

4. The system as described in claim 3, characterized in that, The memory fusion module is used to dynamically retrieve memory information and integrate online and offline data during the dialogue process, ensuring the semantic consistency and behavioral coherence of the subject.

5. The system as described in claim 3, characterized in that, The memory fusion module includes a memory retrieval submodule and an information fusion submodule; (1) Memory retrieval submodule The memory retrieval submodule is used to dynamically retrieve online and offline behavioral memory databases of social subjects to support the generation of current dialogues or behaviors. Specifically, it extracts memory information related to the topic or scenario of the current dialogue from online and offline behavioral memory databases through semantic matching and priority algorithms, and naturally integrates the retrieved memory information into the dialogue. The specific implementation steps are as follows: S21: Determine the retrieval requirements; generate a memory retrieval request based on the topic of the dialogue, the context, or behavioral requirements; S22: Perform memory retrieval; retrieve relevant content from online and offline behavioral memory databases based on semantic matching algorithms; combine keyword matching, semantic embedding vector similarity, and memory priority to filter and sort, and extract valuable memory information; S23: Generate search results; convert the retrieved memory information into natural language expressions that can be embedded in dialogue, so that the generated memory information is naturally connected with the current context; and record the results of this search. (2) Information Fusion Submodule The information fusion submodule integrates the retrieval results from the memory retrieval submodule to generate a unified behavioral context, providing support for subsequent content generation and memory dumping. The specific implementation steps are as follows: S24: Integrate online and offline information, sort the retrieval results of the memory retrieval submodule based on the timeline, and use the semantic disambiguation capability of the large language model to eliminate duplication and conflict; further establish cross-source semantic associations for the sorted information, link online comment behavior with offline participation activities, and form a unified event chain. S25: Generate behavioral context; extract key elements from the event chain and construct a multi-level semantic context.

6. The system as described in claim 5, characterized in that, The information dissemination and memory dump module is used to enable the natural dissemination of information between social subjects, and to perform semantic summarization and long-term storage of important content after the dialogue ends, so as to support the continuation of context in subsequent interactions.

7. The system as described in claim 5, characterized in that, The information dissemination and memory dump module includes an information dissemination submodule and a memory dump submodule; (1) Information dissemination submodule The information dissemination submodule combines information from online and offline behavioral memory databases to generate natural dialogue content that closely resembles real-life contexts, promoting interaction and information diffusion among social participants. The specific implementation steps are as follows: S31: Information filtering; Based on the dialogue context, filter information with communication value from online and offline behavioral memory databases, and use a large language model to evaluate the relevance, interest and scene matching of the information, and prioritize content that is of interest to both parties in the dialogue or has discussion value. S32: Dialogue content generation; The large language model combines the language style and social preferences of the social subjects to personalize the filtering results of the information to be disseminated, that is, to generate language responses that are logically consistent and semantically natural based on the filtering results and the current topic, and then to verify the consistency and grammar of the generated language responses. S33: Information dissemination execution; insert the verified language response content into the multi-level semantic context generated by the memory retrieval and fusion module, initiate and maintain multi-round dialogue, and realize information diffusion and behavioral interaction between social subjects; (2) Memory dump submodule The memory dumping submodule is used to summarize, refine, and store the interaction content in online and offline behavioral memory databases after a dialogue or event ends, using a large language model. This enables dynamic memory accumulation and sharing among social participants. Specific implementation steps include: S34: Generate a dialogue summary; after each round of dialogue, call the large language model to perform semantic analysis, extract key content, and generate a structured summary, adding a timestamp and the identifiers of the participating social entities; S35: Store summary content; store summary content in the online and offline behavioral memory database of the participants in the social interaction, and dynamically adjust memory priority based on the importance, time and frequency of citation; S36: Memory synchronization and optimization; in subsequent interactions, key memories are gradually disseminated to the long-term memories of other social subjects, realizing information sharing and the construction of a social memory network.

8. The system as described in claim 1, characterized in that, This system is applied in the construction of multi-social-agent network systems.

9. A method for simulating social subject memory based on the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: Step 1. Online and offline information collection and storage by social entities Behavioral data and environmental information of social participants are collected from different sources, and a unified structured storage system is established, as follows: Step 1.1) Online information collection by social media participants ① Real-time acquisition of online activity data from different users through social platform interfaces, including browsing history, interaction history, and published content; ② Establish a unified data format specification and storage structure based on the attributes and types of the collected online information, and store it in a memory database; Step 1.2) Offline information collection of social subjects ① Collect data on the daily activities and environment of social subjects, including activity schedules, location awareness, emotional state, and dynamic events; ② The collected offline information is dynamically recorded in a structured or semantic form into the social subject's daily plan and status database for subsequent behavioral decisions, context generation, and information integration; Step 2. Social Subject Memory Retrieval and Information Fusion Step 2.1) Social Subject Memory Retrieval ① Based on the current dialogue topic and the needs of the interaction task or scenario, dynamically retrieve related memory fragments from the social subject memory database; ② By combining historical memory to generate a coherent contextual dialogue flow, the continuity of language and behavior of real users in long-term social interactions is simulated; Step 2.2) Integration of online and offline information for social entities ① Utilize memory retrieval mechanisms to obtain online and offline information related to the current scene or topic; ② Based on the contextual understanding and semantic modeling capabilities of the large language model, multi-source information is semantically fused to form a complete context for dialogue behavior decision-making, providing support for subsequent information dissemination and content generation; Step 3. Information dissemination and memory transfer of social subjects: Step 3.1) Information dissemination by social subjects ① The natural dissemination of information is achieved through dialogue between social subjects. The large language model selects the most valuable content for dissemination and embeds it into the dialogue context based on the dialogue context, semantic association and social motivation. ② Dynamically integrate online and offline information to generate dialogue content that closely reflects real-world situations, thereby achieving authenticity in information dissemination and consistency in the interaction of subject behaviors; Step 3.2) Social Subject Memory Dump ① Summarize and identify key information in the dialogue, including planning, behavioral decisions, and emotional expressions, and transfer it to the long-term memory of the social subject; ② Create time and topic indexes for the memorized content to facilitate subsequent retrieval and recall.

10. The method as described in claim 9, characterized in that, This method is applied in the construction of multi-social agent network systems.

Images