Method and device for optimizing memory ability of large model agent

Abstract

The invention relates to a large model agent memory ability optimization method and device. The method comprises the following steps: acquiring an interaction record; performing content segmentation on the interaction record to obtain a plurality of text segments, and generating a corresponding text index according to each text segment; vectorizing the text segment and the text index, storing the vectorized text segment and the vectorized text index in a vector database, and providing a vector-based retrieval interface; the text segments and the text indexes are stored in a graph database, and a retrieval interface based on graph nodes is provided. By combining the structural advantages of the graph knowledge base and the vector retrieval capability of the RGA, the memory retrieval efficiency and accuracy of the intelligent agent are greatly improved, and more accurate memory calling and information summarization are realized.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and device for optimizing the memory capacity of a large-model intelligent agent. Background Art

[0002] In recent years, with the rapid development of large multimodal model (LMM) technology, artificial intelligence agent (AIAgent) has become an important application form of this technology and has shown broad application prospects in many fields. An agent based on a large model is usually an AI system that understands and performs complex tasks by collecting and processing environmental information, interacting with the surrounding environment, and taking advantage of the large model in planning, tool calling, and reflection.

[0003] The memory function of the agent records the user's questions and the agent's answers during the interaction between the user and the agent. These memories help the agent provide more accurate and detailed responses in complex tasks, long-term planning, and customized scenarios. However, current memory processing and integration technology is still in its early stages, mainly using the following methods:

[0004] (1) Directly integrate historical conversation content: Many LLM agents (such as intelligent customer service or chatbots) directly incorporate historical conversation content into the current interaction context. This method is simple and easy to implement, but the prompt word window length is limited and cannot save a large amount of historical records. In addition, the original conversation record usually contains a lot of redundant information, which affects the efficiency of LLM in generating new replies.

[0005] (2) Summarizing historical conversation content: Summarizing historical conversations and generating summaries through LLM can help the agent understand historical content and generate more targeted responses in a limited context. However, some detailed information may be lost in the process of generating summaries, resulting in the loss of key information.

[0006] (3) Retrieval using RAG (Retrieval Augmented Generation) technology: By dividing the text into paragraphs of appropriate length, the text is converted into vectors using an embedding model and stored in a vector database. During retrieval, the text paragraph closest to the query is found through cosine similarity, thereby achieving efficient semantic retrieval. Compared with direct import or simple summary, RAG can more effectively save and retrieve historical records. However, when dealing with long texts or scenarios with a lot of repetitive content, the retrieval effect of RAG may not be satisfactory.

[0007] Therefore, the performance of existing intelligent agents in terms of memory ability needs to be improved. Summary of the invention

[0008] Based on this, it is necessary to provide a method and device for optimizing the memory capacity of a large model intelligent agent to address the above technical problems.

[0009] In a first aspect, the present application provides a method for optimizing the memory capacity of a large model agent. The method comprises:

[0010] Get interaction records;

[0011] Segment the interaction records to obtain several text segments, and generate corresponding text indexes based on each text segment;

[0012] The text segments and text indexes are vectorized and stored in a vector database, providing a vector-based retrieval interface; the text segments and text indexes are stored in a graph database, providing a graph node-based retrieval interface.

[0013] In one embodiment, the interaction record includes a first record in a cache area and a second record in a storage area;

[0014] The first record provides context support for the current interaction, and the second record is used to update the vector database and the graph database.

[0015] In one embodiment, when the number of interaction pairs in the first record reaches a threshold, the interaction pair first stored in the buffer area is stored in the storage area to become the interaction pair in the second record.

[0016] In one embodiment, the text index includes a multi-level index;

[0017] Generating corresponding text indexes based on each text segment includes:

[0018] Use a large language model to process text segments and generate a primary text index;

[0019] Use the large language model to process the previous level text index to generate the next level text index;

[0020] Summarize the text indexes at each level in order to obtain a multi-level index.

[0021] In one embodiment, the vector-based search interface is used to convert the content to be searched into a vector form, and perform a cosine similarity comparison with the data in the vector database, and obtain the query result based on the comparison result.

[0022] In one embodiment, the graph node-based retrieval interface is used to extract key information according to the content to be retrieved, and jump query in different graph nodes of the graph database for the key information to obtain the query results.

[0023] In a second aspect, the present application also provides a large model intelligent body memory capacity optimization device. The device comprises:

[0024] Memory acquisition module, used to obtain interaction records;

[0025] The memory processing module is used to segment the content of the interaction record, obtain a number of text segments, and generate a corresponding text index according to each text segment;

[0026] The memory retrieval module is used to vectorize text segments and text indexes, store them in a vector database, and provide a vector-based retrieval interface; store text segments and text indexes in a graph database, and provide a graph node-based retrieval interface.

[0027] In a third aspect, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned large model intelligent agent memory capacity optimization method when executing the computer program.

[0028] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in the above-mentioned large-model intelligent agent memory capacity optimization method when executed by a processor.

[0029] In a fifth aspect, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned large model intelligent agent memory capacity optimization method are implemented.

[0030] The above-mentioned large-model intelligent agent memory capacity optimization method and device include the following contents: obtaining interaction records; segmenting the interaction records, obtaining several text segments, and generating corresponding text indexes according to each text segment; vectorizing the text segments and text indexes, and storing them in a vector database, providing a vector-based retrieval interface; storing the text segments and text indexes in a graph database, and providing a graph node-based retrieval interface. Combining the structural advantages of the graph knowledge base and the vector retrieval capability of RGA, the memory retrieval efficiency and accuracy of the intelligent agent are greatly improved, and more accurate memory calls and information summaries are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of a flow chart of a method for optimizing the memory capacity of a large model agent in one embodiment;

[0032] Figure 2 An interactive diagram of a large model agent memory capacity optimization model in one embodiment;

[0033] Figure 3 A schematic diagram of a process for generating a multi-level index in one embodiment;

[0034] Figure 4It is a structural block diagram of a large model intelligent agent memory capacity optimization device in one embodiment. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0036] The present application is an improvement on the intelligent agent based on the big model. Compared with the previous intelligent agent, the biggest change of the intelligent agent based on the big model is that the big model can be used to realize the path planning of the intelligent agent's behavior. The behavior path planning process uses the thinking chain ability of the big model to decompose the user's unclear or complex instructions into steps, and decompose them into clear tool execution or answers. In addition, the big model can be reused to perform new path planning according to the tool execution results or changes in the environment until a suitable path is found.

[0037] The basic module of the agent is responsible for handling the basic interaction needs of users and solving problems by calling tools. This paper uses the Langchain development framework and OpenAI's GPT-4 large model to implement this module. The GPT-4o model is called through the GPT series model calling interface provided in the Langchain framework; and the personality information of the agent is defined through prompt words. The prompt word example is as follows:

[0038] Prompt="You are an AI assistant. You can answer questions people ask you. You can use the following tools to answer questions and you can use your memory to refine your answers. Here are the tools you can use:

[0039] {tools}

[0040] The following is the relevant content you recall: {memory}

[0041] The following is the content of the user's question: {question}

[0042] Please generate a response based on the above content."

[0043] There are three important parameters. Tools is the tool defined for the agent that can be used for calling. In this example, the tool is to obtain web page information through the Apify crawler tool to answer user questions. Question is the information actually entered by the user. The memory content is the main core content of this invention, that is, to filter the content that is most similar to the current session content through memory.

[0044] Based on the above-mentioned large model agent, the embodiment of the present application provides a method for optimizing the memory capacity of a large model agent, such as Figure 1 As shown, the following steps are included:

[0045] Step 102: Obtain interaction records.

[0046] Interaction records are records of interactions between users and agents, including the content of current or recent interactions, as well as the content of historical interactions. These records are the original memory corpus of the agent.

[0047] Step 104 , segment the interaction record into content, obtain a number of text segments, and generate a corresponding text index according to each text segment.

[0048] Since the interaction between users and agents is usually unstructured text and contains a lot of redundant information, the original memory needs to be preprocessed. The preprocessing process mainly includes content segmentation, index generation and vectorization processing.

[0049] Content segmentation is to divide the original chat records into smaller text segments for subsequent processing. The basis for segmentation can be semantic blocks, sentence boundaries or topic transition points.

[0050] Index generation mainly uses language models to generate indexes for segmented text segments. In addition to indexing, the language model also performs semantic understanding on the text segments to generate corresponding summaries.

[0051] Vectorization processing mainly processes the segmented text segments and the generated indexes, using a pre-trained embedding model to convert the text into high-dimensional vectors for storage and fast retrieval.

[0052] Step 106, vectorize the text segment and the text index, and store them in a vector database, providing a vector-based retrieval interface; store the text segment and the text index in a graph database, and provide a graph node-based retrieval interface.

[0053] The processed vector data is stored in the graph database and the vector database. The graph database is responsible for storing structured knowledge graphs, while the vector database stores high-dimensional vectors for the intelligent agent to call in subsequent interactions.

[0054] Graph databases can intuitively represent the interaction between users and agents, while vector databases can handle more complex semantic retrieval tasks. Graph database retrieval responses are relatively slow, but they provide more indirectly related reference information. Vector retrieval responses are relatively fast and can provide directly related reference information. Combining the structural advantages of graph databases and RAG's vector retrieval capabilities, the memory retrieval efficiency and accuracy of agents are greatly improved, achieving more accurate memory calls and information summaries.

[0055] In this embodiment, the vector-based retrieval interface and the graph node-based retrieval interface can be integrated together, and the intelligent agent can generate more targeted and customized responses based on the memory content retrieved by different interfaces, thereby enhancing the intelligence and relevance of the interaction.

[0056] In one embodiment, Figure 2 As shown, the interaction record includes a first record in a cache area and a second record in a storage area; wherein the first record provides context support for the current interaction, and the second record is used to update the vector database and the graph database.

[0057] The cache is used to store short-term memory and save information generated during the current interaction. These contents are closely related to the current interaction and are saved in the cache for immediate retrieval. By setting the size of the cache pool, it can be ensured that only the most relevant recent interaction records are retained. The cached short-term memory is directly provided to the agent to help it understand and optimize the current interaction content.

[0058] The storage area is used to store long-term memories, which record the user's behavior habits, preferences, and earlier interaction history. These contents play an important guiding role in future interactions and can affect the agent's response style and content generation trend. Through regular content organization and optimization, it is kept updated and relevant. In one embodiment, long-term memory can be stored in a graph database.

[0059] Combining the agent's long-term and short-term memory management, it is possible to utilize the current conversation content and historical memory at the same time to improve the agent's memory effect and make it more context-aware.

[0060] In one embodiment, Figure 2 As shown, when the number of interaction pairs in the first record reaches a threshold, the interaction pair first stored in the buffer area is stored in the storage area to become the interaction pair in the second record.

[0061] When the short-term memory buffer overflows, the excess content is collated and stored in long-term memory. This content is periodically collated and filed in the storage area for future interactions.

[0062] The number threshold can be defined as n, and n rounds of interactions can be stored in the cache, where an interaction pair refers to a user asking a question and an agent answering it. When the number of interaction pairs in the cache reaches n, the earliest interaction pair in the current cache is stored in the storage area as long-term memory, that is, it is finally stored in the graph database and vector database in the corresponding form.

[0063] In one embodiment, the text index includes a multi-level index; generating a corresponding text index according to each text segment includes: processing the text segment using a large language model to generate a primary text index; processing the previous level text index using the large language model to generate a next level text index; and aggregating the text indexes at each level in order to obtain a multi-level index.

[0064] like Figure 3 As shown in FIG. 1 , a schematic diagram of a two-level indexing process is shown. The segmented text segments are input into a large language model (LLM) to generate an index corresponding to each text segment. The above index is input into the large language model again to generate a high-level index. The multi-level index adopts a similar structure.

[0065] In other embodiments, the agent also uses prompt words to define the basic settings of the agent. Through specific prompt words, the memory content generated by the large model has stronger structured characteristics, which significantly improves the organization and retrieval efficiency of the memory content and ensures effective application in complex scenarios.

[0066] In one embodiment, the vector-based search interface is used to convert the content to be searched into a vector form, and perform a cosine similarity comparison with the data in the vector database, and obtain the query result based on the comparison result.

[0067] The user input information is directly processed through vectorization to generate vector coordinates, which are compared with the content stored in the vector database for cosine similarity, and the text segment with the highest similarity is returned.

[0068] In one embodiment, the graph node-based retrieval interface is used to extract key information according to the content to be retrieved, and jump query in different graph nodes of the graph database for the key information to obtain the query results.

[0069] The intelligent body will organize the user interaction records and extract the triple relationships therein to generate a graph structure knowledge network. These triples may contain information such as entities, attributes, and relationships, which can better represent the user's behavior and preferences, and store the generated graph structure in the graph database.

[0070] Graph database retrieval can extract key information topics from the content input by the user through a large model, and then query the graph database. Due to the structural advantages of the graph database itself, it is possible to jump queries in different graph nodes based on the key information in the user input to find content with high relevance.

[0071] By searching the nodes and edges in the graph database, the memory content most relevant to the current query is found, and combined with the vector retrieval results, the optimal memory fragment is returned for the agent to use.

[0072] In summary, the present invention proposes a complete set of optimization solutions, covering the generation and management of short-term memory and long-term memory, and comprehensively improves the memory capacity of large model agents by combining graph database, RGB and other technologies. The specific effects are as follows:

[0073] 1. Enhanced recall speed and accuracy: The present invention realizes a faster and more efficient memory recall mechanism by combining long and short-term memory, graph database and RAG technology. The design of the short-term memory cache pool ensures fast access to recent interactive content, while the long-term memory graph database storage ensures accurate retrieval of historical data. Combined with RAG technology, the system can quickly find the most relevant memory fragments and provide accurate and comprehensive recall content in a short time. This multi-level recall capability not only improves the response speed of the intelligent agent, but also greatly improves the accuracy and coverage of the recall.

[0074] 2. Highly abstract and comprehensive memory summarization capabilities: Compared with existing AI agent architectures, especially traditional memory processing and retrieval methods, the present invention can deeply summarize and effectively retrieve memory content with a long time span and large data volume through index generation and multi-level index retrieval technology. The system can generate high-level summaries covering the complete interaction history, allowing the agent to not only provide immediate responses to current tasks, but also provide guidance based on long-term data. In this way, the agent can make more directional suggestions and decisions based on understanding the user's long-term behavior patterns and preferences.

[0075] 3. Memory retrieval capabilities that emphasize both fine and macro aspects: Traditional LLM memory summary methods can often only provide a rough summary of historical records and have difficulty handling fine-grained content recall; while RAG-based retrieval technology can accurately locate specific content, it is difficult to generate an overall summary. The present invention introduces graph database-enhanced memory retrieval technology to enable the intelligent agent to generate an overall summary from a macro level and accurately retrieve specific details when faced with complex historical interactive content. This recall capability that takes into account both the overall and detailed aspects enables the intelligent agent to be flexibly applied in a variety of scenarios to meet the different needs of users.

[0076] In one embodiment, the open source framework GraphRAG is used to implement memory processing. After GraphRAG is deployed, its core prompt words are modified according to needs to enable it to adapt to the function of generating memory retrieval for the intelligent agent; the vectorization processing and RAG retrieval of documents are realized through the open source embedding model jina-embeddings-v2 of JINA company; the open source graph database Node4j is used to build a graph database; the open source vector database Chroma is used to build a vector database.

[0077] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0078] Based on the same inventive concept, the embodiment of the present application also provides a large model intelligent body memory capacity optimization device for implementing the large model intelligent body memory capacity optimization method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more large model intelligent body memory capacity optimization device embodiments provided below can refer to the limitations of the large model intelligent body memory capacity optimization method above, and will not be repeated here.

[0079] In one embodiment, Figure 4 As shown, a large model agent memory capacity optimization device is provided, including: a memory acquisition module 402, a memory processing module 404 and a memory retrieval module 406, wherein:

[0080] Memory acquisition module 402, used to acquire interaction records;

[0081] The memory processing module 404 is used to segment the content of the interaction record, obtain a number of text segments, and generate a corresponding text index according to each text segment;

[0082] The memory retrieval module 406 is used to vectorize the text segments and text indexes, store them in a vector database, and provide a vector-based retrieval interface; store the text segments and text indexes in a graph database, and provide a graph node-based retrieval interface.

[0083] In one embodiment, the interaction record includes a first record in a cache area and a second record in a storage area; wherein the first record provides context support for the current interaction, and the second record is used to update a vector database and a graph database.

[0084] In one embodiment, when the number of interaction pairs in the first record reaches a threshold, the interaction pair first stored in the buffer area is stored in the storage area to become the interaction pair in the second record.

[0085] In one embodiment, the text index includes a multi-level index. The memory processing module 404 is also used to process the text segment using the large language model to generate a primary text index; process the upper level text index using the large language model to generate a lower level text index; and summarize the text indexes at each level in order to obtain a multi-level index.

[0086] In one embodiment, the vector-based search interface is used to convert the content to be searched into a vector form, and perform a cosine similarity comparison with the data in the vector database, and obtain the query result based on the comparison result.

[0087] In one embodiment, the graph node-based retrieval interface is used to extract key information according to the content to be retrieved, and jump query in different graph nodes of the graph database for the key information to obtain the query results.

[0088] Each module in the above-mentioned large-model intelligent body memory capacity optimization device can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0089] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in all the above method embodiments when executing the computer program.

[0090] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0091] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in all the above method embodiments when executed by a processor.

[0092] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0093] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited thereto. The processors involved in each embodiment provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., but are not limited thereto.

[0094] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.

[0095] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for optimizing the memory capacity of a large model agent, characterized in that: The method comprises: Get interaction records; Segment the interaction record to obtain a plurality of text segments, and generate a corresponding text index according to each of the text segments; The text segment and the text index are vectorized and stored in a vector database, and a vector-based retrieval interface is provided; the text segment and the text index are stored in a graph database, and a graph node-based retrieval interface is provided.

2. The method according to claim 1, characterized in that The interaction record includes a first record in a buffer area and a second record in a storage area; The first record provides context support for the current interaction, and the second record is used to update the vector database and the graph database.

3. The method according to claim 2, characterized in that When the number of interaction pairs in the first record reaches a threshold, the interaction pair first stored in the buffer area is stored in the storage area to become the interaction pair in the second record.

4. The method according to claim 1, characterized in that: The text index includes a multi-level index; Generating a corresponding text index according to each of the text segments comprises: Processing the text segment using a large language model to generate a primary text index; Processing the text index of the previous level by using the large language model to generate the text index of the next level; The text indexes at each level are summarized in order to obtain the multi-level index.

5. The method according to claim 1, characterized in that The vector-based search interface is used to convert the content to be searched into a vector form, and perform a cosine similarity comparison with the data in the vector database, and obtain a query result based on the comparison result.

6. The method according to claim 1, characterized in that The graph node-based retrieval interface is used to extract key information according to the content to be retrieved, and to jump query in different graph nodes of the graph database for the key information to obtain query results.

7. A large model intelligent agent memory capacity optimization device, characterized in that: The device comprises: Memory acquisition module, used to obtain interaction records; A memory processing module, used to segment the interaction record, obtain a number of text segments, and generate a corresponding text index according to each of the text segments; The memory retrieval module is used to vectorize the text segment and the text index, store them in a vector database, and provide a vector-based retrieval interface; store the text segment and the text index in a graph database, and provide a graph node-based retrieval interface.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images