Intelligent agent memory data processing method and system and nonvolatile storage medium

Abstract

The invention discloses an agent memory data processing method and system and a nonvolatile storage medium. The method comprises the following steps: determining an interaction text between a target object and a target agent, and determining user intention information corresponding to the interaction text and a memory item extraction result corresponding to the interaction text according to the interaction text; searching a historical memory data set associated with the user intention information and the memory item extraction result from a historical memory database according to the user intention information and the memory item extraction result; determining a comprehensive strength index of each piece of historical memory data in the historical memory data set, and updating the historical memory data according to the comprehensive strength index and the memory item extraction result; and sending update information of the historical memory data to the target agent. The technical problems of low agent response accuracy and poor user experience caused by poor memory data processing capability of the agent are solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and more specifically, to an intelligent body memory data processing method, system and non-volatile storage medium. Background Art

[0002] With the rise of Large Language Model (LLM), various industries have built various intelligent applications around LLM as a basic capability, such as customer service assistants, solution assistants, etc. Many industries have invested in the construction of artificial intelligence services, which enables businesses to be empowered by artificial intelligence to a large extent, greatly enhancing the core competitiveness of enterprises.

[0003] In the process of providing personalized services for thousands of people and iterating related intelligent platforms, large language models have limited memory updating and management capabilities, resulting in poor self-evolution capabilities of the intelligent platform. In subsequent continuous interactions, it is impossible to provide consistent personalized experiences for different users, resulting in a reduced user experience.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present application provide an intelligent agent memory data processing method, system and non-volatile storage medium to at least solve the technical problems of low intelligent agent response accuracy and poor user experience caused by the intelligent agent's poor memory data processing ability.

[0006] According to one aspect of an embodiment of the present application, a method for processing memory data of an intelligent agent is provided, comprising: determining an interaction text between a target object and a target intelligent agent, and determining user intent information corresponding to the interaction text and a memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data for reflecting the semantic information of the interaction text; retrieving a historical memory data set associated with the user intent information and the memory item extraction result from a historical memory database based on the user intent information and the memory item extraction result; determining a comprehensive strength index of each historical memory data in the historical memory data set, and updating the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation relationship with the user intent information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; and sending the update information of the historical memory data to the target intelligent agent.

[0007] Optionally, determining the comprehensive strength index of each historical memory data in the historical memory data set includes: determining the update duration of the historical memory data, and generating a time index of the historical memory data based on the update duration, wherein the size of the time index is negatively correlated with the memory duration of the historical memory data; determining a correlation index of the historical memory data, wherein the correlation index is used to reflect the correlation between the historical memory data and user intention information; determining a preset weight coefficient, wherein the preset weight coefficient is used to determine the weights corresponding to the time index and the correlation index respectively; and determining the comprehensive strength index based on the preset weight coefficient.

[0008] Optionally, generating a time index of historical memory data according to the update duration includes: determining the time index by the following formula: M(t)=M 0 ·e -βt Among them, M 0 is the preset initial time index, e is a natural number, β is the preset time decay constant, and β>0, t is the update duration, and M(t) is the time index.

[0009] Optionally, determining the correlation index of historical memory data includes: determining the prior probability value and the likelihood probability value of the historical memory data, wherein the prior probability value is used to reflect the importance of the historical memory data, and the likelihood probability value is used to reflect the correlation between the historical memory data and the user intention information; determining the correlation index of the historical memory data based on the prior probability value and the likelihood probability value.

[0010] Optionally, updating the historical memory data based on the comprehensive strength index and the memory item extraction results includes: filtering the historical memory data in the historical memory data set based on the comprehensive strength index to obtain the historical memory data set to be processed; determining the relative relationship between the historical memory data to be processed in the historical memory data set to be processed and the memory data in the memory item extraction results, wherein the relative relationship includes that the historical memory data and the memory data are unrelated, the historical memory data and the memory data conflict, and the memory data is a supplementary description of the historical memory data; processing the historical memory data to be processed based on the relative relationship and the memory data.

[0011] Optionally, processing the historical memory data to be processed based on the relative relationship and the memory data includes: when the relative relationship is that the historical memory data and the memory data are unrelated, adding the memory data to the historical memory database; when the relative relationship is that the historical memory data and the memory data conflict, updating the historical memory data based on the memory data; when the relative relationship is that the memory data is a supplementary description of the historical memory data, merging the memory data and the historical memory data.

[0012] Optionally, after retrieving a historical memory data set associated with the user intention information and the memory item extraction result from the historical memory database according to the user intention information and the memory item extraction result, the method further includes: processing the historical memory data set using a mind tree prompting project to obtain multi-layer memory data, wherein the multi-layer memory data is used to indicate the association relationship between the historical memory data set and the user intention information; compressing the historical memory data in the historical memory data set to obtain a compressed historical data memory set; merging the multi-layer memory data and the compressed historical data memory set to obtain a memory data analysis result. Sending the memory data analysis result to the target intelligent agent.

[0013] According to another aspect of the embodiment of the present application, there is also provided an intelligent agent memory data processing system, including: a user intention recognition module, used to determine user intention information corresponding to the interactive text based on the interactive text; a memory extraction module, used to determine the memory item extraction result corresponding to the interactive text based on the interactive text, wherein the memory item extraction result includes memory data used to reflect the semantic information of the interactive text; a memory retrieval module, used to retrieve a historical memory data set associated with the user intention information and the memory item from a historical memory database based on the user intention information and the memory item; a memory update module, used to determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation relationship with the user intention information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; a memory update feedback module, used to send the update information of the historical memory data to the target intelligent agent.

[0014] According to another aspect of an embodiment of the present application, a non-volatile storage medium is provided, in which a program is stored, wherein when the program is running, a device where the non-volatile storage medium is located is controlled to execute the intelligent body memory data processing method.

[0015] According to another aspect of an embodiment of the present application, there is also provided an electronic device, including: a memory and a processor, the processor being used to run a program stored in the memory, wherein the agent memory data processing method is executed when the program is run.

[0016] According to another aspect of an embodiment of the present application, a computer program product is also provided, including a computer program, which implements the intelligent agent memory data processing method when the computer program is executed by a processor.

[0017] In an embodiment of the present application, an interaction text between a target object and a target agent is determined, and user intent information corresponding to the interaction text and a memory item extraction result corresponding to the interaction text are determined based on the interaction text, wherein the memory item extraction result includes memory data for reflecting the semantic information of the interaction text; a historical memory data set associated with the user intent information and the memory item extraction result is retrieved from a historical memory database based on the user intent information and the memory item extraction result; a comprehensive strength index of each historical memory data in the historical memory data set is determined, and the historical memory data is updated based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the association relationship with the user intent information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; a method of sending the update information of the historical memory data to the target agent achieves the purpose of efficiently processing the memory data by identifying the user intent information, retrieving and updating the historical memory data and outputting it to the agent, thereby realizing the technical effect of flexibly responding to user needs, and further solving the technical problems of low agent response accuracy and poor user experience due to the poor memory data processing capability of the agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 It is a structural schematic diagram of a computer terminal (mobile device) provided according to an embodiment of the present application;

[0020] Figure 2 It is a flowchart of an intelligent agent memory data processing method provided according to an embodiment of the present application;

[0021] Figure 3 It is a schematic diagram of a process for obtaining memory data analysis results provided in an embodiment of the present application;

[0022] Figure 4 It is a schematic diagram of a framework of an intelligent body memory data processing system provided according to an embodiment of the present application;

[0023] Figure 5 It is a schematic diagram of the interaction flow between various modules provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0025] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] In order to better understand the embodiments of the present application, the technical terms involved in the embodiments of the present application are explained as follows:

[0027] Large Language Model: Large Language Model (LLM) is a natural language processing model based on deep learning technology that can understand and generate natural language text. Through training with a large amount of data, LLM has the ability to handle complex language tasks and has become a core component of the intelligent agent, supporting its performance in a variety of application scenarios.

[0028] Agent: An agent (AIAgent) is a system that can autonomously perceive the environment, make decisions, and take actions. Agents are usually built on LLM, have the ability to process natural language, and can interact with users in a complex manner. The design of an agent involves multiple components, including sensors, decision-making mechanisms, and actuators, which work together to enable the agent to operate effectively in a dynamic environment.

[0029] Agent cross-session memory: Agent cross-session memory refers to the knowledge and user preferences that can be shared and stored between different sessions in an agent based on LLM. This memory mechanism allows the agent to improve the personalization and accuracy of its responses through historical interactions with users.

[0030] Self-evolving memory updating: Self-evolving memory updating refers to the process whereby the agent autonomously updates and optimizes its memory content through interaction with the environment. This process enables the agent to adapt to a dynamically changing environment, accumulate experience, and improve its decision-making capabilities. For example, the agent can remember past mistakes and avoid repeating the same mistakes in future tasks.

[0031] In the process of providing personalized services for thousands of people and iterating related intelligent agent platforms, large language models have major limitations, such as the inability to use context-based short-term memory efficiently across conversations; the poor generalization ability of personalized content of intelligent agents, which cannot provide different personalized responses to different users in a timely manner; and the limited ability to update and manage memory. This series of problems will ultimately lead to poor self-evolution capabilities of the intelligent agent platform, and may not be able to provide a consistent personalized experience for different users in subsequent continuous interactions. Therefore, it is crucial to introduce a long-term memory processing mechanism, which can effectively improve the intelligence of the model and user experience.

[0032] The main existing methods for implementing intelligent agent memory are as follows.

[0033] Method 1: The agent uses the contextual conversation history as short-term memory. Since large language models have shortcomings in industrial applications, such as the limitation of the number of tokens and the requirement for model reasoning speed, the contextual conversation history is compressed by windowing and memory distillation to avoid problems that arise during the application process. Windowing refers to maintaining the conversation in a certain conversation round in the form of a window and providing it to the LLM as short-term memory; memory distillation refers to summarizing the conversation history within the conversation cycle, thereby reducing tokens and speeding up model reasoning. Both methods have problems such as loss of details and inability to remember across conversations, thus failing to provide users with a continuous and evolving service experience, which will lead to business integration difficulties such as poor business applicability.

[0034] Method 2: The agent integrates business-related knowledge memory through retrieval augmented generation (RAG). Existing agents rely on RAG to perform retrieval augmented generation on business data to provide cross-session business knowledge memory. Within a certain time window, if the business of the relevant enterprise is stable, the RAG processing mode will not cause major problems. However, it cannot be ignored that memory integration in the form of RAG will gradually expose serious business loopholes over time. With too much accumulation of business documents, it is difficult for relevant business personnel to update business documents efficiently and in a timely manner, which will result in the use of old business processing rules at inappropriate times, thereby triggering a series of unnecessary business disputes.

[0035] Method 3: The agent summarizes the context memory and uses it as memory to update old memory or add new memory. Some frameworks extract memory from the conversation context and retrieve the memory through a semantic vectorization model (hereinafter referred to as the Embedding model). This processing method can greatly alleviate the disadvantages of methods 1 and 2, but these frameworks lack the recording of memory metadata and further fine processing, and mainly focus on the implementation of lower-level memory mechanisms such as memory generation, reflection trees, and associative network construction. Since users usually have unclear expressions when using smart applications, if intent recognition processing is not performed, in the process of deploying vectorization services using a general Embedding model, it is usually impossible to accurately obtain relevant memories. In the memory retrieval results, there is no further fine filtering of irrelevant memories, which will introduce greater uncertainty to the relevant services of the agent.

[0036] The above methods have exposed the shortcomings of LLM in the application process when it is the core of intelligent agents from different levels. In order to better provide more personalized and customized intelligent services to users of intelligent applications, there is an urgent need for an effective cognitive persistent memory system for intelligent agents and related self-evolving memory enhancement methods.

[0037] In order to solve the above problems, relevant solutions are provided in the embodiments of the present application, which are described in detail below.

[0038] According to an embodiment of the present application, a method embodiment of an intelligent body memory data processing method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing an intelligent agent memory data processing method. Figure 1As shown, the computer terminal 10 (or mobile device 10) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0040] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0041] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the intelligent body memory data processing method in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned intelligent body memory data processing method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0042] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0043] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0044] In the above operating environment, the embodiment of the present application provides a method for processing intelligent agent memory data, such as Figure 2 As shown, the method comprises the following steps:

[0045] Step S202, determine the interaction text between the target object and the target intelligent agent, and determine the user intention information corresponding to the interaction text and the memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data for reflecting the semantic information of the interaction text.

[0046] Optionally, the agent platform passes the conversation and context (i.e., interactive text) between the user (i.e., the target object) and the agent to the agent memory data processing system through the system's external interface in the form of an API call in an agreed protocol. The protocol content includes, but is not limited to, conversation context, permission group information, user metadata information, and related series configuration information. By collecting multiple types of data, more detailed task information is given to the process. An interface system with a memory change interface and a memory query interface as the core and multiple interfaces as the auxiliary is formed to assist the agent in providing diversified services to the outside world. Each memory persistence evolution and memory query and specification can read necessary data from different data tables through relevant configuration information to build a dynamic memory mapping configuration, thereby meeting different agent configuration situations, greatly enhancing the versatility of the memory system. In order to support cross-session permission group memory sharing, each memory carries the corresponding permission group information. When retrieving, all memories are first screened for permission domains, and then memory retrieval and specification processing are performed on memories with valid access rights.

[0047] When users use natural language to interact with intelligent agents, problems such as unclear semantics, ambiguous references, and ambiguous expressions often occur. For example, using pronouns without clearly referring to the object, incomplete descriptions leading to ambiguous goals, and using jargon or slang that makes it difficult to capture intent. Intention recognition solutions based on large language models (LLMs) are gradually becoming mainstream. These models can learn complex language rules and user behavior patterns from massive data to achieve a more natural and smooth interactive experience. In this embodiment, determining the user intent information corresponding to the interactive text based on the interactive text mainly includes the processing of reference resolution and user demand expansion. This embodiment is based on the Qwen-72B model, and uses the Prompt (prompt word) paradigm of the thinking chain combined with the Few Shot (few sample) technology to construct a Prompt template, thereby activating the powerful language understanding and extrapolation capabilities of Qwen-72B, and placing the user metadata information and dialogue context provided by the application system into the Prompt. By interacting with the model service provider interface, a deep-level reflective analysis result is obtained, and key elements such as entity information and connections between entities are effectively mined, thereby helping the system to better grasp user needs. This embodiment further analyzes the processing results of the model to obtain a user demand question set (i.e., user demand information) after reference resolution, which provides a reliable data premise for the subsequent memory retrieval module and memory convention module, so that the system can be more intelligent and flexible in complex dialogue scenarios and meet diverse user needs. Many long-term memory systems ignore the importance of this module. The lack of this module will seriously lead to the system's inability to effectively understand the user's intentions, which will seriously affect the overall function of the system and user experience.

[0048] Optionally, determining the memory item extraction result corresponding to the interactive text based on the interactive text includes: taking the Qwen-72B model as the core, compiling the corresponding memory extraction prompt according to the business characteristics, such as the intelligent body of the integrated memory system is to provide business assistance to relevant personnel of other businesses, then the need to parse the relevant business memory of the enterprise can be emphasized in the prompt design. This module parses the received dialogue context to obtain a variety of memory tags, including but not limited to memory content, memory available permission group, and memory classification. Further, in order to facilitate the retrieval of memory, this embodiment converts the memory content into a semantic vector (i.e., memory item) through the bge-large-zh model. The model collects the valuable cognitive resources of the user in the context dialogue and continuously injects them into the system, which helps to establish a closer and more fitting interactive experience with the user. The memory extraction in the embodiment helps to timely capture and generalize the user's unique cognitive system throughout the life cycle, which greatly affects the system's availability and personalized service capabilities.

[0049] Optionally, after determining the interactive text between the target object and the target intelligent agent, it also includes sensitive information detection and interception of the conversation (i.e., interactive text) context: through deep learning and strategy in parallel, the context is judged whether to intercept. If the user enters some illegal content, the subsequent memory processing steps are skipped and directly returned to the relevant system. The relevant system prompts the user to enter illegal content or make a reply and other subsequent processes; if the content entered by the user passes the detection mechanism, it enters the memory processing process. Through different application configurations, this patent dynamically loads application configurations, which strengthens the tendency of memory extraction in different applications and greatly enhances the flexibility of the application.

[0050] Specifically, sensitive information detection and interception in conversation contexts include maintaining a library of illegal words for rule matching, using a fine-tuned BERT model for initial semantic matching screening, and using LLM's generalization capabilities and Few Shot learning for deep detection. This multi-layer protection mechanism can effectively filter sensitive topics, insulting language, and false information, while avoiding potential loopholes in rule matching and semantic detection, thereby fully ensuring the security and compliance of content.

[0051] Step S204, based on the user intention information and the memory item extraction result, a historical memory data set associated with the user intention information and the memory item extraction result is retrieved from the historical memory database.

[0052] Optionally, the retrieved historical memory data set will carry complete metadata information (such as the source, time, category, etc. of the historical memory data). The update of historical memory data needs to be based on an efficient memory retrieval mechanism. Retrieval includes building a multi-channel recall mechanism to retrieve the relevant memories obtained by the memory extraction module from a variety of different perspectives, such as semantic similarity and text similarity.

[0053] Specifically, the memory content or the question content for the memory (i.e. the memory item extraction results and user intention information) is vectorized in the memory library, and the memory vector library (i.e. the historical memory database) is searched through the ANN (approximate nearest neighbor) algorithm. The old memory is searched by setting the similarity threshold and combining the top n. When the related memory meets the similarity threshold and the top n requirements at the same time, it is regarded as related memory; the memory content is segmented by the word segmenter, and an inverted index is constructed. After the query content is obtained, the word segmentation is also performed, and the text similarity is judged by the BM25 algorithm. According to actual needs, the system can integrate other search strategies such as knowledge graph similarity and domain ontology similarity to comprehensively expand the breadth and depth of search.

[0054] Optionally, after searching the persisted memory through the retrieval module and retrieving the historical memory data set associated with the user intention information and the memory item extraction result from the historical memory database according to the user intention information and the memory item extraction result, the method further includes: processing the historical memory data set using a mind tree prompting project to obtain multi-layer memory data, wherein the multi-layer memory data is used to indicate the association relationship between the historical memory data set and the user intention information; compressing the historical memory data in the historical memory data set to obtain a compressed historical data memory set; merging the multi-layer memory data and the compressed historical data memory set to obtain a memory data analysis result. Send the memory data analysis result to the target intelligent agent.

[0055] Optionally, in the relevant memory retrieval process, this embodiment clarifies and expands the user's real memory needs through the corresponding user intention understanding module, and further retrieves the relevant memory through the semantic module and the relevant retrieval algorithm. At the same time, due to the limited model capabilities, the retrieved memory may be misaligned with the user's needs. In order to maintain the high availability of memory, this embodiment proposes a memory reduction method to further filter and screen the memory. After the memory reduction processing is completed, the query and reduction results are fed back to the target intelligent agent through the feedback interface.

[0056] Specifically, Figure 3 The flowchart of obtaining the memory data analysis results is shown. The memory convergence protocol module processes the relevant memory (i.e., the historical memory data set) retrieved after being processed by the content intention recognition module. The core processing path is as follows Figure 3 As shown. Since the user intention recognition module retrieves a large number of relevant memories from multiple angles in the process of analyzing the user intention, the historical memory data set may include memories that are irrelevant to the user input, which brings inconvenience in use. For example, the agent may produce errors or hallucinations due to irrelevant memories. In this embodiment, the retrieved memory is deeply filtered and refined, and compressed and extracted by the Qwen-72B model and the LLMlingua algorithm. In order to improve the diversity and effectiveness of memory, the Prompt engineering processing of the thinking tree is used to generate highly concentrated multi-level memories by calling Qwen-72B, that is, it not only contains direct memory content, but also contains related background information, context details and potential meanings. The high-level memory obtained in this way can reveal the deep relationship between different memories and user descriptions, while avoiding the possible information loss of the LLMlingua algorithm. This optimization process helps to remove irrelevant information noise and provide a more accurate and concentrated memory collection, thereby providing high-quality information support for downstream agents or applications.

[0057] Step S206, determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation with the user intention information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point.

[0058] Optionally, the retrieved memory (historical memory data) and related meta-information are semantically analyzed with the extracted new memory, and the effectiveness of memory retrieval is increased by introducing a time decay mathematical model and a Bayesian model. The relevant memory update interface is called for updating, and the memory update interface includes but is not limited to adding new memory, deleting memory, and updating memory. The target time point includes the time point when the interactive text is received or the current time point.

[0059] In the technical solution provided in step S206, determining the comprehensive strength index of each historical memory data in the historical memory data set includes: determining the update duration of the historical memory data, and generating a time index of the historical memory data based on the update duration, wherein the size of the time index is negatively correlated with the memory duration of the historical memory data; determining the correlation index of the historical memory data, wherein the correlation index is used to reflect the correlation relationship between the historical memory data and the user intention information; determining a preset weight coefficient, wherein the preset weight coefficient is used to determine the weights corresponding to the time index and the correlation index respectively; determining the comprehensive strength index based on the preset weight coefficient.

[0060] Optionally, by determining the update duration of historical memory data and generating a time indicator for the historical memory data based on the update duration, it is ensured that the system does not retain outdated information for a long time, thereby maintaining the timeliness of the memory. Furthermore, the relevant memory is controlled through a comprehensive strength indicator, combined with the prompt word project, and the language understanding ability of the large language model is used to perform update judgment processing on the memory.

[0061] Optionally, generating a time index of historical memory data according to the update duration includes: determining the time index by the following formula: M(t)=M 0 ·e -βt Among them, M 0 is the preset initial time index, e is a natural number, β is the preset time decay constant, and β>0, t is the update duration, and M(t) is the time index.

[0062] Specifically, the calculation formula of the time decay function is as follows, where M(t) is the memory strength at the current time t; M 0is the initial strength. This embodiment measures the strength of memory by calculating the frequency of memory recall. β is the time decay constant, and its domain is β>0. t is the interval from the last update to the current retrieval, in days. Under this time decay processing, the memory strength corresponding to the retrieved memory will gradually weaken over time, helping the agent to suppress outdated information, thereby improving the timeliness of memory.

[0063] M(t)=M 0 ·e -βt

[0064] Optionally, determining the correlation index of historical memory data includes: determining the prior probability value and the likelihood probability value of the historical memory data, wherein the prior probability value is used to reflect the importance of the historical memory data, and the likelihood probability value is used to reflect the correlation between the historical memory data and the user intention information; determining the correlation index of the historical memory data based on the prior probability value and the likelihood probability value.

[0065] Specifically, relying solely on the time decay function may cause the system to ignore the actual needs of users, because the decay of memory is based only on time, without considering the user's interaction frequency and contextual information. Therefore, it is possible to mistakenly weaken important memory items that the user is still concerned about, thereby affecting the responsiveness of the agent. To solve this problem, this patent further introduces a Bayesian model to adjust the memory strength by dynamically learning the relevance of user input. The Bayesian update process includes calculating the prior probability P(M i ), likelihood probability P(D|M i ) and the evidence probability P(D), where M i represents the current memory, and D represents the user's input.

[0066] The prior probability is used to reflect the importance of historical memory data, that is, the proportion of the recall times of the historical memory data to the recall times of all retrieved historical memory data. The calculation formula is as follows:

[0067]

[0068] Among them, C Mi Represents the number of recalls of the current memory, C total Represents the statistics of all recall times of this recall memory.

[0069] Specifically, the likelihood probability value is used to reflect the correlation between historical memory data and user intention information. This embodiment uses the correlation obtained by the cross-encoder model to approximate the calculation of the alternative conditional probability to characterize the likelihood probability. Since the dual-encoder model is used in the retrieval step to calculate the vectors of user input and memory respectively, this can greatly improve the processing speed. However, since the input and memory items are independently encoded, this method usually calculates the similarity when the context is complex. It is not accurate enough. For this reason, in order to ensure the effectiveness of memory updating, the likelihood probability is calculated here by the cross-encoder model.

[0070] After determining the prior probability value and likelihood probability value of the historical memory data, multiply the prior probability value and the likelihood probability value and sum them to obtain the value of the evidence probability P(D):

[0071]

[0072] Among them, in order to focus on the memory items directly related to the user input, the efficiency and accuracy of the calculation can be improved. This embodiment only performs a weighted sum of the likelihood probabilities on the retrieved memory data items. Compared with the evaluation statistics for all memory items, this can avoid introducing irrelevant memory items during the calculation, thereby improving the calculation efficiency and accuracy. In addition, this method is more in line with the real-time needs of users and makes the system more responsive. The final Bayesian probability is obtained by fusing the prior probability, likelihood probability, and evidence probability. The fusion formula is as follows:

[0073]

[0074] P(M i |D) is the correlation index of historical memory data.

[0075] The advantage of combining the time decay function with the Bayesian model is that the time decay function effectively manages the timeliness of memory, while the Bayesian model dynamically learns the relevance of user input to achieve more accurate memory updates, thereby improving the system's responsiveness to user needs. This technical solution introduces the hyperparameter α to control the weight ratio between the time decay model and the Bayesian model, and enhances the availability of the entire module by adjusting the operating time of the business system.

[0076] P(M i ′)=α·M(t)+(1-α)·P(M i |D)

[0077] P(M i ′) is the comprehensive strength index.

[0078] In the technical solution provided in step S206, updating the historical memory data based on the comprehensive strength index and the memory item extraction results includes: filtering the historical memory data in the historical memory data set based on the comprehensive strength index to obtain the historical memory data set to be processed; determining the relative relationship between the historical memory data to be processed in the historical memory data set to be processed and the memory data in the memory item extraction results, wherein the relative relationship includes that the historical memory data and the memory data are irrelevant, the historical memory data and the memory data conflict, and the memory data is a supplementary description of the historical memory data; processing the historical memory data to be processed based on the relative relationship and the memory data.

[0079] Optionally, processing the historical memory data to be processed based on the relative relationship and the memory data includes: when the relative relationship is that the historical memory data and the memory data are unrelated, adding the memory data to the historical memory database; when the relative relationship is that the historical memory data and the memory data conflict, updating the historical memory data based on the memory data; when the relative relationship is that the memory data is a supplementary description of the historical memory data, merging the memory data and the historical memory data.

[0080] Specifically, this embodiment establishes a system for knowledge memory iteration. After obtaining the memory extracted from the conversation context, the memory library of the old memory will be searched through the memory retrieval module of the embodiment. Further, after the memory retrieval, the embodiment introduces the time decay model and the Bayesian model to model the retrieved memory to ensure the dynamic and accuracy of memory management. The time decay model calculates and adjusts the memory strength by setting the decay function according to the number of days converted from the timestamp of the memory and the historical recall count statistics, thereby reducing the impact of outdated information on system performance. For example, when the number of recalls between memory A and memory B is the same, and memory B exists longer than memory A, then according to the calculation result of the exponential decay function, the adjusted memory strength of memory B will be weaker than memory A; when the existence time of memory C and memory D is very similar, if the number of recalls of memory C is higher than that of memory D, the memory strength of memory C will be better than that of memory D. This mechanism enables the system to effectively exclude memories that are no longer relevant and improve the fluency and accuracy of the conversation.

[0081] The Bayesian model plays a crucial role in memory updating. By calculating the prior probability, likelihood probability, and evidence probability, the Bayesian model can dynamically evaluate the relationship between the user input D and the recalled memory items M. i Specifically, the cross encoder model is used to calculate the likelihood probability P(D|M i), can quantify the relationship between user input and memory items, thereby guiding the updating and adjustment of memory. For example, if a user inputs feedback about a service, the system will update its likelihood probability by calculating the similarity between the input and the relevant memory items in the memory library. If the input of the new memory is statistically highly correlated with the old memory, its prior probability and likelihood probability will be improved, thereby enhancing the weight of the memory in future decisions. The Bayesian model can not only deal with the conflict between new and old memories, but also provide an adaptive update strategy for the system.

[0082] Processing the historical memory data to be processed based on the relative relationship and memory data includes evaluating the new and old memories. This embodiment is based on the Qwen72B model. The tool call prediction is performed through the Qwen72B model to evaluate the memory iteration from multiple angles. The main angles are memory time proximity, memory content proximity, and memory emotion proximity. If the new memory is irrelevant to the retrieved old memory, the new memory is directly added in the form of addition; if the new memory is inconsistent with the old memory, the old memory is updated. If it conflicts with multiple old memories, the redundant memory is deleted and the latest conflicting memory is selected for memory update; if the new memory is a supplementary description of the old memory, the new and old memories are merged to update the old memory. After obtaining the prediction result of the tool call, the system API is called to perform operations including but not limited to adding, deleting, and updating the content of the memory library. In this way, the system can maximize the potential value of historical memory, promote knowledge iteration, and empower dialogue.

[0083] Suppose a user mentions in a conversation: "I recently used product A and it performed very well." This input will be processed as a new memory. The system first searches for old memory items related to "product A" through the memory retrieval module. Suppose two old memories are retrieved from the memory bank:

[0084] Memory 1: "The user mentioned the price of product A last month.";

[0085] Memory 2: "The user mentioned his preference for product B three months ago.";

[0086] The system calculates the prior probability, likelihood probability, and evidence probability:

[0087] Prior probability P(M 1 ) and P(M 2 ) is based on historical data statistics;

[0088] Likelihood probability P(D|M 1 ) and P(D|M 2 ) uses natural language processing methods to calculate the relevance of user input and retrieved old memories through a cross-encoder model.

[0089] The system applies a time decay model to decay the memory items. For each old memory item, the memory strength M(t) is calculated as: M(t) = M 0 ·e -βt , we get M 1 (t) and M 2 (t).

[0090] When confirming the valid memory subset, the system combines the results of the Bayesian model with the memory strength after time decay. Specifically, the system combines the Bayesian update with the time decay through the weight α to form a comprehensive memory strength P(M i ′). Furthermore, low memory strength is filtered through the threshold, and only high comprehensive strength memory is processed. The relationship between memories is comprehensively considered through the Qwen72B model. If the comprehensive memory strength of the new memory and memory 1 and the various large model extrapolation evaluation indicators are higher than memory 2, the system will update the weight of memory 1, reflecting the user's recognition of the performance of product A. If the system detects a contradiction between the new memory and the old memory (for example, the user once expressed dissatisfaction with product A), the old memory will be updated through the priority strategy to retain the most relevant information.

[0091] Step S208, sending the update information of the historical memory data to the target agent.

[0092] Through the above steps, the processing efficiency of the cognitive persistence of the intelligent agent and the effect of enhanced cognitive self-evolution can be achieved, and the problems of poor short-term memory migration, poor personalized generalization ability of the intelligent agent, lack of systematic business memory update and management capabilities, and poor memory refinement management capabilities when LLM is used as the core of the intelligent agent can be solved.

[0093] The embodiment of the present application builds a highly robust, highly available, self-evolving, and extrapolable cognitive long-term memory system, which improves the efficiency, accuracy, and business availability of LLM in complex scenarios, thereby further enhancing the overall performance and practical value of the intelligent agent system, enabling the intelligent agent to more effectively respond to complex business needs in practical applications, while reducing resource consumption and improving user experience. Specifically, the embodiment of the present application has the following advantages:

[0094] 1. This patent proposes a highly available content specification method: a memory specification method is proposed to further filter and optimize the memory to improve the high availability of retrieval, ensure the high quality and availability of memory data, provide accurate and centralized memory collection, and provide high-quality information support for downstream applications.

[0095] 2. This patent introduces a time decay model and a Bayesian model into the memory update module. Under the evaluation of the mathematical model, the strength of memory is corrected, and the availability of memory and the effectiveness of retrieval are balanced, so that the memory update module of the intelligent body can adapt to dynamic changes and filter out outdated information, maintaining a flexible response to user needs. It can be extrapolated to the relevant retrieval mechanism as a core function, dynamically adjusting the memory processing strategy according to different applications, and enhancing the adaptability and flexibility of the system.

[0096] 3. This patent takes deep learning technology as the core and focuses on strengthening user usage optimization: taking deep learning technology as the core, integrating business strategies, and accurately analyzing the intentions and semantic understanding of the dialogue context between users and intelligent agents. Through deep learning technology, accurate analysis of user intentions is performed, which improves the dialogue system's ability to understand the context, can timely capture the user's unique cognitive system, and enhance personalized service capabilities.

[0097] 4. This patent can realize cross-session permission group memory sharing, improving system flexibility and versatility.

[0098] An embodiment of the present application provides an intelligent agent memory data processing system, including: a user intention recognition module, used to determine user intention information corresponding to the interactive text based on the interactive text; a memory extraction module, used to determine a memory item extraction result corresponding to the interactive text based on the interactive text, wherein the memory item extraction result includes memory data used to reflect the semantic information of the interactive text; a memory retrieval module, used to retrieve a historical memory data set associated with the user intention information and the memory item from a historical memory database based on the user intention information and the memory item; a memory update module, used to determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation with the user intention information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; a memory update feedback module, used to send the update information of the historical memory data to the target intelligent agent.

[0099] Optionally, Figure 4 shows the framework diagram of the agent memory data processing system, Figure 5 The interaction process between each module is shown, such as Figure 4 and Figure 5As shown, the intelligent agent memory data processing system is mainly composed of six core modules: user intention recognition module, memory extraction module, memory retrieval module, memory update module, memory specification module, and memory processing feedback module. The context detection and interception module is an optional module for functional expansion, supplemented by business strategies, to analyze the conversation context between the user and the intelligent agent, detect the context content, and extract, retrieve, specify and update the memory. In order for other systems to better access related functions, the system proposed in this patent also establishes a supporting access plan and processing feedback mechanism.

[0100] After the intelligent agent receives the user's interactive text, on the one hand, it intercepts illegal information through the context detection and interception module, extracts the intercepted data from the memory, and after retrieval, it is handed over to the memory update module for processing, and the update result is fed back to the intelligent agent. On the other hand, the data is retrieved and transmitted to the reduction module. The reduction module further compresses and extracts information from the historical memory data set, determines a more accurate historical memory data subset, and transmits it to the intelligent agent for interaction with the user.

[0101] An embodiment of the present application provides a non-volatile storage medium, in which a program is stored, wherein when the program is running, the device where the non-volatile storage medium is located is controlled to execute the following agent memory data processing method: determine the interaction text between the target object and the target agent, and determine the user intention information corresponding to the interaction text and the memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data used to reflect the semantic information of the interaction text; retrieve a historical memory data set associated with the user intention information and the memory item extraction result from a historical memory database based on the user intention information and the memory item extraction result; determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the association relationship with the user intention information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; send the update information of the historical memory data to the target agent.

[0102] An embodiment of the present application provides an electronic device, comprising: a memory and a processor, the processor being used to run a program stored in the memory, wherein the following agent memory data processing method is executed when the program is running: determining an interaction text between a target object and a target agent, and determining user intent information corresponding to the interaction text and a memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data used to reflect the semantic information of the interaction text; retrieving a historical memory data set associated with the user intent information and the memory item extraction result from a historical memory database based on the user intent information and the memory item extraction result; determining a comprehensive strength index of each historical memory data in the historical memory data set, and updating the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation relationship with the user intent information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; and sending the update information of the historical memory data to the target agent.

[0103] An embodiment of the present application provides a computer program product, including a computer program, which implements the following agent memory data processing method when executed by a processor: determining the interaction text between the target object and the target agent, and determining the user intention information corresponding to the interaction text and the memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data used to reflect the semantic information of the interaction text; retrieving a historical memory data set associated with the user intention information and the memory item extraction result from a historical memory database based on the user intention information and the memory item extraction result; determining the comprehensive strength index of each historical memory data in the historical memory data set, and updating the historical memory data based on the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update duration of the historical memory data and the correlation relationship with the user intention information, and the update duration is the duration between the most recent update time point of the historical memory data and the target time point; sending the update information of the historical memory data to the target agent.

[0104] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0105] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0107] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0108] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the relevant technology or all or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc. Various media that can store program codes.

[0109] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for processing agent memory data, characterized in that: include: Determine an interaction text between a target object and a target agent, and determine user intention information corresponding to the interaction text and a memory item extraction result corresponding to the interaction text based on the interaction text, wherein the memory item extraction result includes memory data for reflecting semantic information of the interaction text; Retrieving a historical memory data set associated with the user intention information and the memory item extraction result from a historical memory database according to the user intention information and the memory item extraction result; Determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data according to the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update time of the historical memory data and the correlation relationship with the user intention information, and the update time is the time between the most recent update time point of the historical memory data and the target time point; Sending update information of the historical memory data to the target agent.

2. The agent memory data processing method according to claim 1, characterized in that: Determining the comprehensive strength index of each piece of historical memory data in the historical memory data set includes: Determine the update duration of the historical memory data, and generate a time index of the historical memory data according to the update duration, wherein the size of the time index is negatively correlated with the memory duration of the historical memory data; Determining a correlation index of the historical memory data, wherein the correlation index is used to reflect the correlation relationship between the historical memory data and the user intention information; Determining a preset weight coefficient, wherein the preset weight coefficient is used to determine the weights corresponding to the time indicator and the relevance indicator respectively; The comprehensive strength index is determined according to the preset weight coefficient.

3. The agent memory data processing method according to claim 2, characterized in that: The time index for generating the historical memory data according to the update duration includes: The time index is determined by the following formula: M(t)=M0·e -βt Wherein, M0 is the preset initial time index, e is a natural number, β is the preset time decay constant, and β>0, t is the update duration, and M(t) is the time index.

4. The agent memory data processing method according to claim 2, characterized in that: The correlation indexes for determining the historical memory data include: Determine a priori probability value and a likelihood probability value of the historical memory data, wherein the priori probability value is used to reflect the importance of the historical memory data, and the likelihood probability value is used to reflect the correlation between the historical memory data and the user intention information; The correlation index of the historical memory data is determined according to the prior probability value and the likelihood probability value.

5. The agent memory data processing method according to claim 1, characterized in that: The updating process of the historical memory data according to the comprehensive strength index and the memory item extraction result includes: Filtering the historical memory data in the historical memory data set according to the comprehensive strength index to obtain a historical memory data set to be processed; Determine the relative relationship between the historical memory data to be processed in the historical memory data set to be processed and the memory data in the memory item extraction result, wherein the relative relationship includes that the historical memory data is irrelevant to the memory data, the historical memory data conflicts with the memory data, and the memory data is a supplementary description of the historical memory data; The historical memory data to be processed is processed according to the relative relationship and the memory data.

6. The agent memory data processing method according to claim 5, characterized in that: Processing the to-be-processed historical memory data according to the relative relationship and the memory data includes: In the case where the relative relationship is that the historical memory data and the memory data are irrelevant, adding the memory data to the historical memory database; When the relative relationship is that the historical memory data and the memory data conflict, updating the historical memory data according to the memory data; In the case where the relative relationship is that the memory data is a supplementary description of the historical memory data, the memory data and the historical memory data are merged.

7. The agent memory data processing method according to claim 1, characterized in that: After retrieving a historical memory data set associated with the user intention information and the memory item extraction result from a historical memory database according to the user intention information and the memory item extraction result, the method further includes: Processing the historical memory data set using a mind tree prompting project to obtain multi-layer memory data, wherein the multi-layer memory data is used to indicate the association relationship between the historical memory data set and the user intention information; Compressing the historical memory data in the historical memory data set to obtain a compressed historical data memory set; Merging the multi-layer memory data and the compressed historical data memory set to obtain a memory data analysis result; Send the memory data analysis result to the target agent.

8. An agent memory data processing system, characterized in that: include: A user intention recognition module, used to determine user intention information corresponding to the interactive text based on the interactive text; A memory extraction module, used to determine a memory item extraction result corresponding to the interactive text according to the interactive text, wherein the memory item extraction result includes memory data for reflecting the semantic information of the interactive text; A memory retrieval module, used to retrieve a historical memory data set associated with the user intention information and the memory item from a historical memory database according to the user intention information and the memory item; A memory update module, used to determine the comprehensive strength index of each historical memory data in the historical memory data set, and update the historical memory data according to the comprehensive strength index and the memory item extraction result, wherein the comprehensive strength index is used to reflect the update time of the historical memory data and the correlation relationship with the user intention information, and the update time is the time between the most recent update time point of the historical memory data and the target time point; The memory update feedback module is used to send the update information of the historical memory data to the target agent.

9. A non-volatile storage medium, characterized in that: The non-volatile storage medium stores a program, wherein when the program is running, the device where the non-volatile storage medium is located is controlled to execute the intelligent body memory data processing method described in any one of claims 1 to 7.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program, when running, executes the intelligent body memory data processing method described in any one of claims 1 to 7.

11. A computer program product, characterized in that It comprises a computer program which, when executed by a processor, implements the agent memory data processing method according to any one of claims 1 to 7.

Images