An agent long-term memory dynamic management method and system based on evidence governance

Abstract

The application discloses an agent long-term memory dynamic management method and system based on evidence governance, comprising: acquiring current input information and a short-term context memory window, and judging a memory operation type; generating to-be-processed memory information when performing memory writing or updating; determining to create a new memory node or incorporate an existing memory node according to a new memory declaration; updating a current belief, a confidence degree and a valid time interval, and maintaining an aggregated memory node and an isolated memory node; when performing memory retrieval, performing hierarchical retrieval, rearrangement and confidence degree screening on the aggregated memory node and the isolated memory node to obtain a target memory node set, inputting the target memory node set, a current user question and the short-term context memory window into a large language model agent to generate an agent reply. The application can improve the traceability, updating accuracy and retrieval efficiency of long-term memory, and reduce the influence of expired memory or conflict memory on the agent reply.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, natural language processing, agent long-term memory management and trusted retrieval technology, and in particular to an evidence-based method and system for dynamic management of agent long-term memory. Background Technology

[0002] In recent years, with the development of large language models and intelligent agent technology, intelligent agents have been widely used in scenarios such as personal assistants, office automation, medical question answering, legal consultation, scientific research assistance, and long-term task execution. Unlike traditional single-turn question answering systems, intelligent agents need to retain information such as user facts, preferences, task status, historical decisions, and environmental experiences during continuous interaction. Therefore, long-term memory capability has become an important foundation for intelligent agents to provide continuous services and make personalized decisions.

[0003] For example, Chinese patent document CN120688545A discloses a method for modeling long-term memory of an intelligent agent based on a memory network, including: storing vector data in a vector database and storing structured metadata in a relational database; receiving user input requests, performing vector similarity retrieval in the vector database to obtain semantically similar segments; performing structured data query in the relational database to obtain structured metadata; retrieving memory summaries; forming candidate datasets; using the filtered results as memory information, inputting the memory information into a large language model through prompt word engineering to generate response content; combining the user input request and the generated response content to form a new interaction record; and processing the new interaction record and the historical interaction records stored in the vector database and relational database through the large language model to generate a memory summary.

[0004] Chinese patent document CN121413649A discloses a method for long-term memory of a robot bionic intelligent agent, including: acquiring dialogue text and context information through real-time interaction; generating a structured memory summary by calling a large language model based on the dialogue record; persistently storing the memory summary in a relational database when the session ends; and automatically associating the user's historical memory as dialogue context input in subsequent interactions, so that the large language model can provide personalized responses based on long-term memory.

[0005] However, existing long-term memory methods for intelligent agents mostly rely on text summarization, vector databases, or knowledge graphs to store historical information. They typically focus more on memory storage and semantic retrieval, lacking systematic management of evidence regarding the source of memories, conflicting evidence, temporal changes, and credibility. This makes it difficult to determine whether a memory is still valid and suitable for use in a specific task. Especially in high-risk scenarios such as healthcare, law, and finance, erroneous or expired memories can lead to serious consequences. The system needs not only to remember the content but also to clearly define the basis of the memory, its validity period, conflict situations, and usage risks. Furthermore, existing memory structures often rely on fixed types or simple similarity relationships, making it difficult to adapt to the complex and ever-changing user information and long-term state changes in open scenarios.

[0006] Therefore, there is an urgent need for a unified management method and system that can write, update, retrieve, and delete long-term memories of intelligent agents based on evidentiary relationships, so as to improve the traceability, timeliness, reliability, and security of long-term memories. Summary of the Invention

[0007] In view of the problems of untraceable source, unclear conflict updates, low retrieval efficiency and difficulty in managing redundant memory in existing agent long-term memory, this invention provides a dynamic management method and system for agent long-term memory based on evidence governance. Through memory nodes, evidence chains, aggregated memory nodes and hierarchical retrieval mechanisms, it can improve the traceability, update accuracy and retrieval efficiency of long-term memory, and reduce the impact of expired or conflicting memory on agent responses.

[0008] A dynamic management method for long-term memory of intelligent agents based on evidence governance includes the following steps: S1: Obtain the current input information and short-term context memory window received by the agent during the interaction process, input them into the large language model agent, determine the memory operation type corresponding to the current input information, and obtain the memory decision; S2: When the memory decision instruction performs memory writing or memory update, the current input information is used as the extraction object, and the short-term context memory window is used as the auxiliary context to extract new memory statements, current beliefs, new evidence and evidence generation time as memory information to be processed; S3: Based on the new memory declarations in the memory information to be processed, perform similarity matching and rearrangement on existing aggregated memory nodes and isolated memory nodes to determine whether to create new memory nodes or merge them into existing memory nodes; S4: When there are memory nodes to be merged, determine the type of evidence relationship corresponding to the new evidence based on the memory information to be processed and the current belief of the memory node to be merged, write the evidence chain corresponding to the memory node to be merged, and update the current belief, confidence level and effective time interval to obtain the updated memory node; if the current belief is updated to the opposite of the original belief, then swap the contents of the support chain and the conflict chain. S5: For newly created or updated memory nodes, maintain aggregated memory nodes and isolated memory nodes based on the newly created or updated memory nodes; S6: When the memory decision instruction performs memory retrieval, perform hierarchical retrieval, rearrangement and confidence filtering on aggregated memory nodes and isolated memory nodes to obtain the target memory node set; and input the current user question, the target memory node set and the short-term context memory window into the large language model agent to generate the agent's response.

[0009] Furthermore, in S3 and subsequent steps, the memory node includes the memory statement, current belief, confidence level, supporting evidence chain, conflicting evidence chain, time-updated evidence chain, valid time interval, call statistics, creation time, and update time.

[0010] Furthermore, in step S1, the memory decision includes not performing a memory operation, writing a memory, updating a memory, and retrieving a memory.

[0011] Furthermore, step S3 specifically includes: Based on the new memory declaration in the memory information to be processed, perform first-level similarity matching on existing aggregated memory nodes and existing isolated memory nodes; when the matching result includes aggregated memory nodes, perform second-level similarity matching on the child memory nodes under the aggregated memory nodes; generate a set of similar nodes based on the isolated memory nodes obtained from the first-level similarity matching and the child memory nodes obtained from the second-level similarity matching. The new memory declaration and all memory nodes in the set of similar nodes are input into the rearrangement model and sorted to obtain the sequence of similar nodes; Based on the sequence of similar nodes, determine whether to merge the memory information to be processed into an existing memory node or to create a new memory node; when the highest rearrangement score is less than the preset merging threshold, create a new memory node based on the memory information to be processed; when the highest rearrangement score is greater than or equal to the preset merging threshold, determine the memory node corresponding to the highest rearrangement score as the memory node to be merged.

[0012] Furthermore, in step S4, the newly created memory node is represented as follows: ; in, A declaration for new memories, For the current belief, As the initial confidence level, This indicates that the initial evidence chain is supported. Since the node is newly created, the conflict evidence chain and the time evidence chain are empty. express; This is the initial valid time interval. For initial call statistics, The time when the evidence was generated.

[0013] Further, in step S4, the type of evidence relationship corresponding to the new evidence is determined, including supporting relationships, conflicting relationships, and time update relationships; the evidence chain corresponding to the memory node to be merged is written, according to the following rules: Support relationships: ; Conflict Relationships: ; Time update relationship: ; in, These are the supporting evidence chain, conflict evidence chain, and time update evidence chain for the memory nodes to be merged.

[0014] Furthermore, in step S4, the confidence update is represented as follows: ; in, , , The value is 1 when the evidentiary relationship is a supporting relationship, a conflicting relationship, or a time update relationship, respectively; otherwise, the value is 0. This is the corresponding adjustment coefficient.

[0015] When the evidentiary relationship is a time-updated relationship, update the current belief and the effective time interval; when the updated current belief is contrary to the previous current belief, swap the attribution of evidence in the supporting evidence chain and the conflicting evidence chain before the update.

[0016] Furthermore, the specific process of step S5 is as follows: For newly created or updated memory nodes, if the similarity between the memory node and an existing aggregated memory node is greater than or equal to a preset aggregation threshold, the memory node is added to the corresponding aggregated memory node; if there is no existing aggregated memory node that meets the preset aggregation threshold, and the similarity between the memory node and at least one isolated memory node is greater than or equal to a preset clustering threshold, a new aggregated memory node is generated based on the memory node and the isolated memory node; otherwise, the memory node is saved as an isolated memory node.

[0017] Furthermore, step S6 specifically includes: When the memory decision instruction executes memory retrieval, a first-level search is performed on aggregated memory nodes and isolated memory nodes in the memory set based on the current user question; when the first-level search hits an aggregated memory node, a second-level search is performed on the memory nodes contained in the aggregated memory node; the memory nodes obtained from the second-level search are merged with the isolated memory nodes hit by the first-level search, and filtered according to the current time and the effective time interval of each memory node to obtain a set of relevant memory nodes; The relevant memory node set is rearranged to obtain the rearranged memory node sequence; the target memory node set is obtained by filtering the memory nodes in the rearranged memory node sequence based on their confidence scores; the current user question, the target memory node set, and the short-term context memory window are input into the large language model agent to generate the agent's response.

[0018] An intelligent agent long-term memory dynamic management system based on evidence governance includes a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned intelligent agent long-term memory dynamic management method.

[0019] Compared with the prior art, the present invention has the following beneficial effects: By recording the source and relationship of evidence through the chain of evidence, the traceability of memories is improved; by updating the current belief, confidence level and effective time interval, the probability of expired or conflicting memories being mistakenly recalled is reduced; by aggregating memory nodes, secondary retrieval and rearrangement filtering, the efficiency and relevance of long-term memory retrieval are improved; and by performing decay management through memory retention scores, long-term memory redundancy is reduced. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A schematic diagram of the process of the intelligent agent long-term memory dynamic management method based on evidence governance provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the module structure of the intelligent agent long-term memory dynamic management system based on evidence governance provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0024] like Figure 1As shown, a dynamic management method for long-term memory of intelligent agents based on evidence governance includes the following steps: S101: Obtain the current input information and short-term context memory window, determine the memory operation type corresponding to the current input information according to the preset memory trigger judgment criteria, and obtain the memory decision.

[0025] S102: When the memory decision instruction performs memory writing or memory update, the current input information is used as the extraction object, and the short-term context memory window is used as the auxiliary context to extract new memory statements, current beliefs, new evidence and evidence generation time as memory information to be processed.

[0026] S103: Based on the new memory declaration, perform first-level similarity matching on existing aggregated memory nodes and existing isolated memory nodes; when the first-level matching result includes aggregated memory nodes, perform second-level similarity matching on their child memory nodes; generate a set of similar nodes based on the matching results, and obtain a sequence of similar nodes through a rearrangement model.

[0027] S104: Determine the write object based on the highest rearrangement score in the similar node sequence; create a new memory node when the highest rearrangement score is less than the preset merging threshold, otherwise determine the corresponding memory node as the memory node to be merged.

[0028] S105: When there is a memory node to be merged, determine the type of evidence relationship corresponding to the new evidence, write the new evidence into the corresponding evidence chain, and update the current belief, confidence level and effective time interval to obtain the updated memory node; when the current belief changes to a belief opposite to the current belief before the update, swap the evidence attribution in the supporting evidence chain and the conflicting evidence chain before the update.

[0029] S106: Update the aggregated memory node according to the newly created or updated memory node; add the corresponding aggregated memory node when the memory node meets the preset aggregation threshold, generate a new aggregated memory node when it meets the preset clustering threshold with at least one isolated memory node, otherwise save it as an isolated memory node.

[0030] S107: When the memory decision instruction executes memory retrieval, a first-level retrieval is performed on aggregated memory nodes and isolated memory nodes based on the current user question; when the first-level retrieval hits an aggregated memory node, a second-level retrieval is performed on its child memory nodes; the memory nodes obtained from the second-level retrieval and the isolated memory nodes hit by the first-level retrieval are merged, and a set of relevant memory nodes is obtained by filtering according to the effective time interval.

[0031] S108: Rearrange the relevant memory node set and filter it according to the confidence level to obtain the target memory node set; input the current user question, the target memory node set and the short-term context memory window into the large language model agent to generate the agent's response.

[0032] Furthermore, the memory node is represented as follows: ; ; in, For the purpose of remembering the statement, For the current belief, For confidence level, These are the supporting chain of evidence, the conflicting chain of evidence, and the time-updated chain of evidence. For the valid time interval, To access statistics, For the number of calls, The most recent call time, and These represent the creation time and the update time, respectively.

[0033] Furthermore, the memory information to be processed is represented as follows: ; ; in, For memory information to be processed, A declaration for new memories, For the current belief, As new evidence, For the time when the evidence was generated, This serves as a source of evidence.

[0034] Furthermore, in step S104, the write target is determined based on the sequence of similar nodes: ; ; ; in, For the highest rearrangement score, For memory nodes to be merged, The preset merging threshold is used. Newly created memory nodes are represented as follows: ; in, As the initial confidence level, This indicates that the initial chain of evidence is supported. This is the initial valid time interval. Statistics for the initial call.

[0035] Furthermore, in step S5, the types of evidentiary relationships include supporting relationships, conflicting relationships, and time-update relationships, represented as follows: ; The system adds new evidence to the corresponding chain of evidence based on the type of evidence relationship: ; ; ; in, , These represent support relationships, conflict relationships, and time update relationships, respectively. These are the supporting evidence chain, conflict evidence chain, and time update evidence chain for the memory nodes to be merged.

[0036] Furthermore, updates to current beliefs, confidence levels, and effective time intervals include: ; in, The value is 1 when the evidentiary relationship is a supporting relationship, a conflicting relationship, or a time update relationship, respectively; otherwise, the value is 0. This is the corresponding adjustment coefficient.

[0037] When the evidentiary relationship is a time-updated relationship: ; When the updated current belief is contrary to the previous current belief, the attribution of evidence in the supporting chain and conflicting chain of evidence is updated as follows: ; Furthermore, the aggregated memory node is represented as follows: ; in, For the aggregation node identifier, For aggregation declaration, For the set of member memory node identifiers, This represents the aggregate confidence level.

[0038] Furthermore, in steps S107 and S108, the system performs a first-level search on aggregated memory nodes and isolated memory nodes; when the first-level search hits an aggregated memory node, a second-level search is performed on its sub-memory nodes; the search results are filtered according to the current time, valid time interval, rearrangement score and confidence level to obtain the target memory node set and generate an agent response.

[0039] Furthermore, the method of the present invention also includes memory decay and deletion methods: ; ; ; ; in, Duration of inactivity Duration of existence This is the last update time. Retain fractions for memory.

[0040] ; in, This indicates the memory decay processing action determined based on the memory retention score, with values ​​including retention, compression, archiving, and deletion. These are the retention threshold, the archiving threshold, and the deletion threshold, respectively.

[0041] Based on the same inventive principle, this invention also provides an intelligent agent long-term memory dynamic management system based on evidence governance, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned intelligent agent long-term memory dynamic management method.

[0042] Based on the same inventive principle, embodiments of the present invention also provide a dynamic management system for long-term intelligent agent memory based on evidence governance, such as... Figure 2 As shown, it includes: The system comprises the following modules: a memory decision module for outputting memory decisions; a memory extraction module for generating memory information to be processed; a similarity matching module for generating a set of similar nodes; a rearrangement judgment module for generating a sequence of similar nodes and determining whether to merge or create new memory nodes; an evidence chain update module for writing the evidence chain and updating the current belief, confidence level, and effective time interval; an aggregation management module for updating aggregated memory nodes and isolated memory nodes; a memory retrieval module for generating a set of related memory nodes; a response generation module for generating agent responses; a memory decay module for performing retention, compression, archiving, or deletion; and a storage module for saving memory nodes, aggregated memory nodes, and intermediate results from each stage.

[0043] The following example of a psychological counseling-assisted intelligent agent will be used to verify the effectiveness of the embodiments of the present invention.

[0044] The psychologist or psychological counseling assistant agent engages in multiple rounds of dialogue with the patient. The system needs to continuously record the patient's sleep status, sources of stress, emotional changes, coping strategies, and high-risk expressions. In subsequent dialogues, the system decides whether to use the memory, use the memory cautiously, or trigger further confirmation based on the confidence level of the memory, the effective time interval, and the chain of evidence.

[0045] The method in this invention includes memory decision-making, memory extraction, similarity matching, rearrangement judgment, evidence chain update, aggregation management, memory retrieval, response generation, and memory decay management.

[0046] In one example of psychological counseling, the patient inputs: "I've been sleeping very poorly for the past two weeks, often waking up at 3 a.m.." The system retrieves the current input information. and Short Context Memory Window Output memory decision Then extract the memory information to be processed. ,in The patient has recent sleep problems. The patient has had poor sleep for the past two weeks. Input for this round of users, This is the dialogue time.

[0047] The system will Match with existing aggregated memory nodes and isolated memory nodes. If a "sleep state" aggregated memory node is matched, continue matching within its child memory nodes to obtain a set of similar nodes. The similar node sequence is obtained after rearrangement. If the highest rearrangement score is less than the preset merging threshold, a new memory node is created; otherwise, the memory node corresponding to the highest rearrangement score is determined as the memory node to be merged.

[0048] When the patient subsequently inputs, "The doctor told me to take the medication for the next few days, and my sleep has improved somewhat since yesterday," the system outputs a memory decision. The input is then used as new evidence. If the evidence relationship type is time-update relationship, it is written into the time-update evidence chain, the current belief is updated to "the patient's sleep has improved after taking the medication", and the effective time interval is updated; if the updated current belief is the opposite of the previous one, the supporting evidence chain and the conflicting evidence chain are re-divided according to the updated current belief.

[0049] When a patient asks, "How has my sleep been lately?", the system outputs a memory decision. First, aggregated memory nodes and isolated memory nodes are searched. If a "sleep state" aggregated memory node is found, a secondary search is performed on its child memory nodes. The system merges the search results and filters the target memory node set based on the valid time interval and confidence level. And generate an intelligent agent response. .

[0050] During long-term operation, the system calculates the memory retention score based on the confidence level, number of calls, most recent call time, creation time, and update time of each memory node, and then... Perform operations such as retaining, compressing, archiving, or deleting memory nodes.

[0051] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic management of long-term memory of intelligent agents based on evidence governance, characterized in that, Includes the following steps: S1: Obtain the current input information and short-term context memory window received by the agent during the interaction process, input them into the large language model agent, determine the memory operation type corresponding to the current input information, and obtain the memory decision; S2: When the memory decision instruction performs memory writing or memory update, the current input information is used as the extraction object, and the short-term context memory window is used as the auxiliary context to extract new memory statements, current beliefs, new evidence and evidence generation time as memory information to be processed; S3: Based on the new memory declarations in the memory information to be processed, perform similarity matching and rearrangement on existing aggregated memory nodes and isolated memory nodes to determine whether to create new memory nodes or merge them into existing memory nodes; S4: When there are memory nodes to be merged, determine the type of evidence relationship corresponding to the new evidence based on the memory information to be processed and the current belief of the memory node to be merged, write the evidence chain corresponding to the memory node to be merged, and update the current belief, confidence level and effective time interval to obtain the updated memory node; if the current belief is updated to the opposite of the original belief, then swap the contents of the support chain and the conflict chain. S5: For newly created or updated memory nodes, maintain aggregated memory nodes and isolated memory nodes based on the newly created or updated memory nodes; S6: When the memory decision instruction performs memory retrieval, hierarchical retrieval, rearrangement and confidence filtering are performed on aggregated memory nodes and isolated memory nodes to obtain the target memory node set; The current user question, the target memory node set, and the short-term context memory window are then input into the large language model agent to generate the agent's response.

2. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, In S3 and subsequent steps, memory nodes include memory statements, current beliefs, confidence levels, supporting evidence chains, conflicting evidence chains, time-updated evidence chains, valid time intervals, call statistics, creation time, and update time.

3. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, In step S1, the memory decision includes not performing a memory operation, writing a memory, updating a memory, and retrieving a memory.

4. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, Step S3 specifically includes: Based on the new memory declaration in the memory information to be processed, perform first-level similarity matching on existing aggregated memory nodes and existing isolated memory nodes; when the matching result includes aggregated memory nodes, perform second-level similarity matching on the child memory nodes under the aggregated memory nodes; generate a set of similar nodes based on the isolated memory nodes obtained from the first-level similarity matching and the child memory nodes obtained from the second-level similarity matching. The new memory declaration and all memory nodes in the set of similar nodes are input into the rearrangement model and sorted to obtain the sequence of similar nodes; Based on the sequence of similar nodes, determine whether to merge the memory information to be processed into an existing memory node or to create a new memory node; when the highest rearrangement score is less than the preset merging threshold, create a new memory node based on the memory information to be processed; when the highest rearrangement score is greater than or equal to the preset merging threshold, determine the memory node corresponding to the highest rearrangement score as the memory node to be merged.

5. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, In step S4, the newly created memory node is represented as follows: ； in, A declaration for new memories, For the current belief, As the initial confidence level, This indicates the initial support for the evidence chain; since the node is newly created, the conflict evidence chain and the time evidence chain are empty, so [the following is used]. express; This is the initial valid time interval. For initial call statistics, The time when the evidence was generated.

6. The method for dynamic management of agent long-term memory based on evidence governance according to claim 5, characterized in that, In step S4, the type of evidence relationship corresponding to the new evidence is determined, including supporting relationship, conflict relationship, and time update relationship; Write the evidence chain corresponding to the memory node to be merged, according to the following rules: Support relationships: ; Conflict Relationships: ; Time update relationship: ; in, These are the supporting evidence chain, conflict evidence chain, and time update evidence chain for the memory nodes to be merged.

7. The method for dynamic management of agent long-term memory based on evidence governance according to claim 6, characterized in that, In step S4, the confidence update is represented as follows: ； in, , , The value is 1 when the evidentiary relationship is a supporting relationship, a conflicting relationship, or a time update relationship, respectively; otherwise, the value is 0. For the corresponding adjustment coefficient; When the evidentiary relationship is a time-updated relationship, update the current belief and the effective time interval; when the updated current belief is contrary to the previous current belief, swap the attribution of evidence in the supporting evidence chain and the conflicting evidence chain before the update.

8. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, The specific process of step S5 is as follows: For newly created or updated memory nodes, if the similarity between the memory node and an existing aggregated memory node is greater than or equal to a preset aggregation threshold, the memory node is added to the corresponding aggregated memory node; if there is no existing aggregated memory node that meets the preset aggregation threshold, and the similarity between the memory node and at least one isolated memory node is greater than or equal to a preset clustering threshold, a new aggregated memory node is generated based on the memory node and the isolated memory node; otherwise, the memory node is saved as an isolated memory node.

9. The method for dynamic management of agent long-term memory based on evidence governance according to claim 1, characterized in that, Step S6 specifically includes: When the memory decision instruction executes memory retrieval, a first-level search is performed on aggregated memory nodes and isolated memory nodes in the memory set based on the current user question; when the first-level search hits an aggregated memory node, a second-level search is performed on the memory nodes contained in the aggregated memory node; the memory nodes obtained from the second-level search are merged with the isolated memory nodes hit by the first-level search, and filtered according to the current time and the effective time interval of each memory node to obtain a set of relevant memory nodes; The relevant memory node set is rearranged to obtain the rearranged memory node sequence; the target memory node set is obtained by filtering the memory nodes in the rearranged memory node sequence based on their confidence scores; the current user question, the target memory node set, and the short-term context memory window are input into the large language model agent to generate the agent's response.

10. A dynamic management system for long-term memory of intelligent agents based on evidence governance, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the intelligent agent long-term memory dynamic management method according to any one of claims 1-9.

Images