Multi-agent memory conflict arbitration method based on version vector and policy library

Abstract

The application provides a multi-agent memory conflict arbitration method based on a version vector and a policy library, comprising the following sub-steps: S1: generating a memory version vector; S2: performing conflict detection and classification; S3: performing arbitration strategy matching based on the memory conflict type; and S4: performing feedback learning; the method introduces a memory version vector and an association index, realizes cross-agent and cross-context memory sharing, and breaks information islands; the memory conflict type is identified based on a semantic similarity score, a time sequence difference and a task correlation feature, and arbitration strategies are matched based on the memory conflict type, effective information is completely retained, and the reliability is improved through confidence judgment; the model and the strategy weight are optimized through feedback learning, autonomous evolution is realized, repeated conflicts are reduced, and system overhead is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a multi-agent memory conflict arbitration method based on version vectors and policy libraries. Background Technology

[0002] In multi-agent collaborative systems, agents need to frequently share and store memory information to achieve task coordination, information transmission, and context persistence. Memory serves as the accumulation of agent knowledge, the basis for decision-making, and the foundation for collaboration. Its consistency, accuracy, and availability directly determine the task reliability and collaboration efficiency of the multi-agent collaborative system. As multi-agent systems are widely used in complex scenarios, the scale of memory continues to expand and its sources become more diverse. Memory conflicts between different agents regarding the same object, event, or state have become commonplace. How to efficiently and intelligently handle memory conflicts has become a key issue restricting the implementation of multi-agent systems. However, the existing technology has the following drawbacks: Independent memory storage leads to information silos: Each intelligent agent independently maintains its own memory, lacking a unified and standardized memory sharing and association mechanism. It is unable to perceive the knowledge accumulation and updates of other intelligent agents, resulting in scattered information that is difficult to reuse and low efficiency in cross-agent collaboration. The conflict resolution method is simple, but it loses valuable information: When multiple memory entities have different memories of the same information, the existing technical solutions only use a single strategy of overwriting timestamps or ignoring them directly. This makes it impossible to determine the nature of the conflict, and valuable conflict information is discarded. Only a single conclusion is retained, which lacks credibility. Without self-learning and experience accumulation mechanisms, conflicts recur: Each conflict is an independent event, and no experience is accumulated in solving it, leading to the recurrence of the same type of conflict. The system cannot learn and optimize from historical conflicts, and each conflict needs to be reprocessed, which significantly increases the system's computation and scheduling overhead. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a multi-agent memory conflict arbitration method based on version vectors and policy libraries.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: This invention proposes a multi-agent memory conflict arbitration system based on version vectors and a policy library. The system includes a memory version vector generation module, a conflict detection and classification module, a policy matching and arbitration module, and a feedback learning module. The memory version vector generation module is used to generate memory version vectors, establish associated indexes, and form memory version chains. The conflict detection and classification module is used to filter candidate memories to form a candidate conflict set, construct a conflict feature vector, and determine the type of memory conflict. The strategy matching and arbitration module selects an arbitration strategy based on the memory conflict type and the success rate weight of the strategy library, and generates a conflict handling log. The feedback learning module verifies the arbitration effect, collects error samples, and updates the policy weights.

[0005] This invention also proposes a multi-agent memory conflict arbitration method based on version vectors and policy libraries, comprising the following sub-steps: S1: Generate a memorized version vector; Pre-set trigger conditions: The memory bank is monitored and monitored in real time through the memory bank monitoring component. When the creation or update of any memory in the memory bank is detected, the memory version vector generation module is triggered, and the memory version vector generation module generates the corresponding memory version vector for that memory. The memory version vector includes the source agent identifier, generation timestamp, confidence score, context fingerprint, and task type; The source agent identifier is the unique agent identifier that generated the memory, which is read from the memory by the memory version vector generation module. The generated timestamp is the time recorded by the memory version vector generation module when the memory is created or updated; The confidence score is the score of the memory version vector output by the memory version vector generation module, and the value range is a floating-point number between [0, 1]. The context fingerprint is a unique hash identifier generated by the memory version vector generation module based on the content in the session context environment where the agent is located; the agent is the agent corresponding to the source agent identifier; The task type is a task classification label divided and labeled by the memory version vector generation module according to the agent's session context environment, including data analysis, text processing, decision suggestion, etc. Establish an association index between the memory version vector and the corresponding memory, and save it to the system's data storage area; When a memory is updated, the historical versions of the memory are preserved, and each version is assigned a version number and a timestamp, forming a memory version chain.

[0006] S2: Perform conflict detection and classification; S21: Select candidate memories and construct a candidate conflict set; When the memory bank monitoring component detects a new write operation in the memory bank, that is, when a new memory is added to the memory bank, it triggers the memory version vector generation module to generate the memory version vector corresponding to the new memory and simultaneously triggers the conflict detection and classification module. The conflict monitoring and classification module reads the text content of the new memory and calls the built-in semantic encoding model to perform semantic parsing and vector encoding on the text content, generating a semantic vector corresponding to the new memory; The similarity between the semantic vector of a new memory and the semantic vector of any memory in the memory bank is calculated one by one using a vector retrieval algorithm. Memories with similarity higher than a preset similarity threshold are selected as candidate memories. All selected candidate memories are saved into a set to form a candidate conflict set. S22: Obtain multiple memory pairs and calculate the features of each memory pair; The new memory is paired with each candidate memory in the candidate conflict set to obtain multiple memory pairs; the conflict monitoring and classification module calculates the features of each memory pair; the features include semantic similarity score, temporal difference, and task relevance; S23: Determine the type of memory conflict for each memory pair; The semantic similarity score, temporal difference, and task relevance of each memory pair are concatenated according to a preset feature order to form the conflict feature vector of each memory pair. The conflict feature vector is input into a pre-trained conflict type classifier. The conflict type classifier performs feature operations and category discrimination on the input conflict feature vector and outputs the memory conflict type corresponding to the conflict feature vector. The memory conflict type includes factual contradiction, semantic redundancy, complementary perspective, and timeliness difference. Specifically, a baseline feature vector is pre-configured for each of the four types of memory conflict. The similarity algorithm is used to calculate the vector similarity between the conflict feature vector and the four baseline feature vectors. The numerical values ​​of the four vector similarities are compared, and the memory conflict type to which the baseline feature vector with the maximum vector similarity belongs is determined as the memory conflict type corresponding to the current conflict feature vector. At the same time, the maximum value of the four vector similarities is used as the confidence score output by the conflict type classifier. Furthermore, if the obtained confidence level is lower than the preset confidence level threshold, the current conflict feature vector is marked as manually reviewed, and the memory conflict type is not determined for the time being. The memory conflict type will be manually confirmed and labeled later.

[0007] S3: Matching arbitration strategies based on memory conflict types; The strategy matching and arbitration module matches the corresponding arbitration strategy from the pre-built strategy library based on the memory conflict type obtained in step S2 and executes the arbitration operation. The strategy library contains the association mapping relationship between memory conflict types and various arbitration strategies, and stores the success rate weight of various arbitration strategies for different memory conflict types. Based on the association mapping relationship, the module selects the optional arbitration strategies under the current memory conflict type, sorts them according to the success rate weight of each arbitration strategy, selects the arbitration strategy with the highest success rate weight as the final execution strategy, and executes the arbitration operation. After the arbitration operation is performed, arbitration information is added to the memory version vector corresponding to each memory involved in the conflict. The arbitration information includes arbitration strategy, arbitration basis, arbitration time, arbitration executor, and arbitration result, and a conflict resolution log is generated. After each memory conflict arbitration is completed and a conflict resolution log is generated, the conflict resolution log is archived and stored as a historical arbitration case.

[0008] The arbitration basis is the memory conflict type, the arbitration time is the timestamp of the execution of the arbitration strategy, the arbitration executor includes automatic system confirmation and manual confirmation, and the arbitration result includes the memory finally selected in this memory conflict, the memory conflict type determination result, etc.

[0009] S4: Provide feedback and learn; S41: Generate a list of memories; Once the conflict resolution log is generated and the arbitration-related process is completed, the feedback learning module is triggered; the memory of this conflict is extracted from the generated conflict resolution log and a memory list is generated. S42: Verify the arbitration effect and generate feedback learning samples; The trigger feedback learning module obtains the task execution results output by the multi-agent collaborative system after executing business tasks through the interactive interface. If the task execution results are normal and meet expectations, the arbitration operation corresponding to this memory conflict is marked as correct; otherwise, the corresponding arbitration operation is marked as pending verification. For arbitration operations marked as pending verification, manual review or feedback learning modules analyze the reasons for business task anomalies to determine whether they are caused by arbitration operations. If the arbitration operation is determined to be the cause, the feedback learning module will label the corresponding conflict feature vector, error handling result, and correct handling result together and add them to the training database as feedback learning samples; the error result is the arbitration result in the conflict handling log; the correct handling result is the correct memory version and correct memory conflict type corresponding to this memory conflict obtained after manual review and re-deduction by the system.

[0010] S43: Update the success rate weights of each arbitration strategy in the strategy library; The success rate of each arbitration strategy in the statistical strategy library under different memory conflict types is calculated. If the success rate of a certain arbitration strategy under a certain memory conflict type is lower than the preset threshold, the success rate weight of the arbitration strategy is reduced, and the weight of other candidate arbitration strategies is increased accordingly. Historical arbitration cases are analyzed in batches according to a preset cycle to generate conflict resolution effectiveness reports.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This method introduces a memory version vector mechanism to assign each memory a multi-dimensional label such as source agent identifier, context fingerprint, and task type, and establishes a unified association index to achieve cross-agent and cross-context memory association and efficient sharing, break down information silos, and improve information reuse rate and collaborative efficiency. By identifying the type of memory conflict and matching the corresponding arbitration strategy, conflicted memories are not directly deleted. Based on the type, methods such as coexistence of primary and reference memories, semantic fusion, retention of multiple versions, and simultaneous retention of new and old memories are adopted to fully preserve all valuable conflict information. At the same time, confidence-weighted judgment is used instead of simple timestamp judgment, which avoids the loss of effective information and improves the reliability of memory selection. This method constructs a conflict type classifier based on semantic similarity, temporal difference, and task relevance to distinguish four types of conflicts. It accurately judges the conflict by combining contextual fingerprints and task type, and sets a confidence threshold to enable manual review, thereby improving the accuracy and pertinence of conflict identification and processing. Incremental optimization of the conflict classifier is achieved by incorporating erroneous arbitration cases into the training library, while dynamically updating the success rate weights of each strategy in the strategy library. This enables autonomous learning, avoids repeated conflicts, reduces system overhead, and continuously improves conflict handling capabilities. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the steps of the multi-agent memory conflict arbitration method based on version vectors and policy libraries proposed in this invention. Detailed Implementation

[0013] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0014] This invention proposes a multi-agent memory conflict arbitration system based on version vectors and a policy library. The system includes a memory version vector generation module, a conflict detection and classification module, a policy matching and arbitration module, and a feedback learning module. The memory version vector generation module is used to generate memory version vectors, establish associated indexes, and form memory version chains. The conflict detection and classification module is used to filter candidate memories to form a candidate conflict set, construct a conflict feature vector, and determine the type of memory conflict. The strategy matching and arbitration module selects an arbitration strategy based on the memory conflict type and the success rate weight of the strategy library, and generates a conflict handling log. The feedback learning module verifies the arbitration effect, collects error samples, and updates the policy weights.

[0015] like Figure 1 As shown, this invention also proposes a multi-agent memory conflict arbitration method based on version vectors and a policy library, including the following sub-steps: S1: Generate a memorized version vector; Pre-set trigger conditions: The memory bank is monitored and monitored in real time through the memory bank monitoring component. When the creation or update of any memory in the memory bank is detected, the memory version vector generation module is triggered, and the memory version vector generation module generates the corresponding memory version vector for that memory. The memory refers to information units stored in a memory bank and generated by an intelligent agent; The memory version vector includes the source agent identifier, generation timestamp, confidence score, context fingerprint, and task type; The source agent identifier is the unique agent identifier that generated the memory, which is read from the memory by the memory version vector generation module. The generated timestamp is the time recorded by the memory version vector generation module when the memory is created or updated; The confidence score is the score of the memory version vector output by the memory version vector generation module, and the value range is a floating-point number between [0, 1]. The context fingerprint is a unique hash identifier generated by the memory version vector generation module based on the content in the session context environment where the agent is located; the agent is the agent corresponding to the source agent identifier; The task type is a task classification label divided and labeled by the memory version vector generation module according to the agent's session context environment, including data analysis, text processing, decision suggestion, etc. Establish an association index between the memory version vector and the corresponding memory, and save it to the system's data storage area; When a memory is updated, the historical versions of the memory are preserved, and each version is assigned a version number and a timestamp, forming a memory version chain.

[0016] Furthermore, each memory corresponds to multiple versions, and each version of the memory corresponds to a memory version vector.

[0017] S2: Perform conflict detection and classification; S21: Select candidate memories and construct a candidate conflict set; When the memory bank monitoring component detects a new write operation in the memory bank, that is, when a new memory is added to the memory bank, it triggers the memory version vector generation module to generate the memory version vector corresponding to the new memory and simultaneously triggers the conflict detection and classification module. The conflict monitoring and classification module reads the text content of the new memory and calls the built-in semantic encoding model to perform semantic parsing and vector encoding on the text content, generating a semantic vector corresponding to the new memory; The similarity between the semantic vector of a new memory and the semantic vector of any memory in the memory bank is calculated one by one using a vector retrieval algorithm. Memories with similarity higher than a preset similarity threshold are selected as candidate memories. All selected candidate memories are saved into a set to form a candidate conflict set. S22: Obtain multiple memory pairs and calculate the features of each memory pair; The new memory is paired with each candidate memory in the candidate conflict set to obtain multiple memory pairs; the conflict monitoring and classification module calculates the features of each memory pair; the features include semantic similarity score, temporal difference, and task relevance; Specifically, semantic similarity score: The cosine similarity between the semantic vectors corresponding to the two memories in each memory pair is calculated using the cosine similarity algorithm, and the resulting cosine similarity value is the semantic similarity score. Temporal difference: Based on the association index, the generation timestamps of the memory version vectors corresponding to the two memories in each memory pair are extracted, and the difference between the generation timestamps corresponding to the two memories is the temporal difference; Task relevance: Based on the association index, extract the context fingerprint and task type in the memory version vector of the two memories in each memory pair; compare the context fingerprint and task type of the two memories respectively, and assign values ​​according to the preset assignment rules to obtain the assignment results corresponding to the context fingerprint and task type respectively; In this embodiment, when the context fingerprint matching is consistent, a value of 1 is assigned and when the matching is inconsistent, a value of 0 is assigned. When the task type matching is consistent, a value of 1 is assigned and when the matching is inconsistent, a value of 0 is assigned. The two assigned values ​​are weighted and summed based on preset weight coefficients, and then divided by the sum of weight coefficients to complete the weighted average operation, thus obtaining the task relevance of the memory pair. S23: Determine the type of memory conflict for each memory pair; The semantic similarity score, temporal difference, and task relevance of each memory pair are concatenated according to a preset feature order to form the conflict feature vector of each memory pair. The conflict feature vector is input into a pre-trained conflict type classifier. The conflict type classifier performs feature operations and category discrimination on the input conflict feature vector and outputs the memory conflict type corresponding to the conflict feature vector. The memory conflict type includes factual contradiction, semantic redundancy, complementary perspective, and timeliness difference. Furthermore, the conflict type classifier is trained in advance using supervised learning models such as random forest, SVM or neural network. The training data comes from the labeled data of historical conflict handling cases, and incremental training is performed periodically using the new data generated by feedback learning. Specifically, a baseline feature vector is pre-configured for each of the four types of memory conflict. The similarity algorithm is used to calculate the vector similarity between the conflict feature vector and the four baseline feature vectors. The numerical values ​​of the four vector similarities are compared, and the memory conflict type to which the baseline feature vector with the maximum vector similarity belongs is determined as the memory conflict type corresponding to the current conflict feature vector. At the same time, the maximum value of the four vector similarities is used as the confidence score output by the conflict type classifier. Furthermore, if the obtained confidence level is lower than the preset confidence level threshold, the current conflict feature vector is marked as manually reviewed, and the memory conflict type is not determined for the time being. The memory conflict type will be manually confirmed and labeled later.

[0018] S3: Matching arbitration strategies based on memory conflict types; The strategy matching and arbitration module matches the corresponding arbitration strategy from the pre-built strategy library based on the memory conflict type obtained in step S2 and executes the arbitration operation. The strategy library contains the association mapping relationship between memory conflict types and various arbitration strategies, and stores the success rate weight of various arbitration strategies for different memory conflict types. Based on the association mapping relationship, the module selects the optional arbitration strategies under the current memory conflict type, sorts them according to the success rate weight of each arbitration strategy, selects the arbitration strategy with the highest success rate weight as the final execution strategy, and executes the arbitration operation. The specific arbitration procedure is as follows: When the memory conflict type is factual contradiction, the voting strategy with the highest success rate weight under this memory conflict type is selected; the confidence scores of all versions of the memory version vectors corresponding to the two memories in the memory pair are statistically analyzed, the version with the highest confidence score is determined as the master memory, the other versions are retained as reference memories, and the memory conflict type is marked in each version of the memory; When the memory conflict type is semantic redundancy, the merging strategy with the highest success rate weight under this memory conflict type is selected; the text content of the two memories in the memory pair is semantically deduplicated, non-repetitive information is extracted and fused to generate a comprehensive memory, the confidence score of the comprehensive memory is the average confidence score of the two memories in the memory pair, the context fingerprint of the comprehensive memory is the union of the context fingerprints of the two memories, and the task type of the comprehensive memory is the union of the task types of the two memories. When the memory conflict type is complementary perspective, select the strategy with the highest success rate weight under this memory conflict type and retain multiple versions; at the same time, retain two memories in the memory pair, mark the two memories as complementary and add a "substitute" label to each memory, and then select the corresponding version of the memory according to the current situation. When the memory conflict type is time difference, select the uncertain labeling strategy with the highest success rate weight under this memory conflict type; extract the generation timestamps of the memory version vectors corresponding to the two memories in the memory pair, retain the memory with the later generation timestamp as the master memory, and add a "may be outdated" label to the memory with the earlier generation timestamp and retain it; After the arbitration operation is performed, arbitration information is added to the memory version vector corresponding to each memory involved in the conflict. The arbitration information includes arbitration strategy, arbitration basis, arbitration time, arbitration executor, and arbitration result, and a conflict resolution log is generated. After each memory conflict arbitration is completed and a conflict resolution log is generated, the conflict resolution log is archived and stored as a historical arbitration case.

[0019] The arbitration basis is the memory conflict type, the arbitration time is the timestamp of the execution of the arbitration strategy; the arbitration executor includes both automatic system confirmation and manual confirmation; the arbitration result includes the memory finally selected in this memory conflict, the memory conflict type determination result, etc. If the system automatically determines the type of memory conflict and executes the arbitration strategy, the arbitrator is marked as "system automatic"; if a human participates in determining the type of memory conflict or confirming the arbitration strategy, the arbitrator is marked as "human confirmation".

[0020] S4: Provide feedback and learn; S41: Generate a list of memories; Once the conflict resolution log is generated and the arbitration-related process is completed, the feedback learning module is triggered; the memory of this conflict is extracted from the generated conflict resolution log and a memory list is generated. S42: Verify the arbitration effect and generate feedback learning samples; The trigger feedback learning module obtains the task execution results output by the multi-agent collaborative system after executing business tasks through the interactive interface. If the task execution results are normal and meet expectations, the arbitration operation corresponding to this memory conflict is marked as correct; otherwise, the corresponding arbitration operation is marked as pending verification. For arbitration operations marked as pending verification, manual review or feedback learning modules analyze the reasons for business task anomalies to determine whether they are caused by arbitration operations. Specifically, if any of the following situations exist—incorrect identification of memory conflict type, incorrect matching of arbitration strategy, inappropriate selection of master memory version, or incorrect confidence score—and the business task anomaly occurs in the business process after using the memory from this arbitration, then the task anomaly is determined to be caused by this arbitration operation; otherwise, the task anomaly is determined to be unrelated to this arbitration operation. If the arbitration operation is determined to be the cause, the feedback learning module will label the corresponding conflict feature vector, error handling result, and correct handling result together and add them to the training database as feedback learning samples; the error result is the arbitration result in the conflict handling log; the correct handling result is the correct memory version and correct memory conflict type corresponding to this memory conflict obtained after manual review and re-deduction by the system.

[0021] S43: Update the success rate weights of each arbitration strategy in the strategy library; The success rate of each arbitration strategy in the statistical strategy library under different memory conflict types is calculated. If the success rate of a certain arbitration strategy under a certain memory conflict type is lower than the preset threshold, the success rate weight of the arbitration strategy is reduced, and the weight of other candidate arbitration strategies is increased accordingly. Historical arbitration cases are analyzed in batches according to a preset cycle to generate conflict resolution effectiveness reports.

[0022] The present invention has been described by the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A multi-agent memory conflict arbitration method based on version vectors and policy libraries, characterized in that: Includes the following steps: S1: Generate a memorized version vector; When the creation or update of any memory in the memory bank is detected, the memory version vector generation module generates a corresponding memory version vector for that memory. The memory version vector includes the source agent identifier, generation timestamp, confidence score, context fingerprint, and task type; Establish an association index between the memory version vector and the corresponding memory, and save it to the system's data storage area; S2: Perform conflict detection and classification; S21: Select candidate memories and construct a candidate conflict set; When a new memory is detected being added to the memory bank, the memory version vector generation module is triggered to generate the memory version vector corresponding to the new memory and the conflict detection and classification module is triggered simultaneously. The conflict monitoring and classification module generates a semantic vector corresponding to the new memory, and filters out candidate memories corresponding to the new memory from the memory bank. The candidate memories are saved into a set to form a candidate conflict set. S22: Obtain multiple memory pairs and calculate the features of each memory pair; The new memory is paired with each candidate memory in the candidate conflict set to obtain multiple memory pairs; the conflict monitoring and classification module calculates the features of each memory pair. The features include semantic similarity score, temporal difference, and task relevance; S23: Determine the type of memory conflict for each memory pair; The semantic similarity score, temporal difference, and task relevance of each memory pair are concatenated according to a preset feature order to form the conflict feature vector of each memory pair. The conflict type classifier outputs the conflict feature vectors and the corresponding memory conflict types. S3: Matching arbitration strategies based on memory conflict types; The strategy matching and arbitration module matches the corresponding arbitration strategy from the pre-built strategy library based on the memory conflict type and executes the arbitration operation; After the arbitration operation is performed, arbitration information is added to the memory version vector corresponding to each memory involved in the conflict, and a conflict resolution log is generated. S4: Provide feedback and learn; S41: Generate a list of memories; S42: Verify the arbitration effect and generate feedback learning samples; S43: Update the success rate weights of each arbitration strategy in the strategy library.

2. The multi-agent memory conflict arbitration method based on version vectors and policy libraries as described in claim 1, characterized in that: In step S1, the memory bank is monitored and monitored in real time through the memory bank monitoring component. When the creation or update of any memory in the memory bank is detected, the memory version vector generation module is triggered, and the memory version vector generation module generates the corresponding memory version vector for that memory. The source agent identifier is the unique agent identifier that generated the memory, which is read from the memory by the memory version vector generation module. The generated timestamp is the time recorded by the memory version vector generation module when the memory is created or updated; The confidence score is the score of the memory version vector output by the memory version vector generation module, and the value range is a floating-point number between [0, 1]. The context fingerprint is a unique hash identifier generated by the memory version vector generation module based on the content in the session context environment where the agent is located; the agent is the agent corresponding to the source agent identifier; The task type is a task classification label divided and labeled by the memory version vector generation module according to the agent's session context environment, including data analysis, text processing, and decision suggestion; When a memory is updated, the historical versions of the memory are preserved, and each version is assigned a version number and a timestamp to form a memory version chain. Each memory corresponds to multiple versions, and each version of the memory corresponds to a memory version vector.

3. The multi-agent memory conflict arbitration method based on version vectors and policy libraries as described in claim 1, characterized in that: The specific details of step S2 are as follows: S21: Select candidate memories and construct a candidate conflict set; When the memory bank monitoring component detects a new write operation in the memory bank, that is, when a new memory is added to the memory bank, it triggers the memory version vector generation module to generate the memory version vector corresponding to the new memory and simultaneously triggers the conflict detection and classification module. The conflict monitoring and classification module reads the text content of the new memory and calls the built-in semantic encoding model to perform semantic parsing and vector encoding on the text content, generating a semantic vector corresponding to the new memory; The similarity between the semantic vector of a new memory and the semantic vector of any memory in the memory bank is calculated one by one using a vector retrieval algorithm. Memories with similarity higher than a preset similarity threshold are selected as candidate memories. All selected candidate memories are saved into a set to form a candidate conflict set. S22: Obtain multiple memory pairs and calculate the features of each memory pair; Semantic similarity score: The cosine similarity between the semantic vectors corresponding to the two memories in each memory pair is calculated using the cosine similarity algorithm. The resulting cosine similarity value is the semantic similarity score. Temporal difference: Based on the association index, the generation timestamps of the memory version vectors corresponding to the two memories in each memory pair are extracted, and the difference between the generation timestamps corresponding to the two memories is the temporal difference; Task relevance: Based on the association index, extract the context fingerprint and task type in the memory version vector of the two memories in each memory pair; compare the context fingerprint and task type of the two memories respectively, and assign values ​​according to the preset assignment rules to obtain the assignment results corresponding to the context fingerprint and task type respectively; The two assignment results are weighted and summed based on preset weight coefficients, and then divided by the sum of weight coefficients to complete the weighted average calculation, thus obtaining the task relevance of the memory pair. S23: Determine the type of memory conflict for each memory pair; The conflict feature vector is input into a pre-trained conflict type classifier. The conflict type classifier performs feature operations and category discrimination on the input conflict feature vector and outputs the memory conflict type corresponding to the conflict feature vector. The memory conflict type includes factual contradiction, semantic redundancy, complementary perspective, and timeliness difference. Pre-configure corresponding baseline feature vectors for four types of memory conflict. Calculate the vector similarity between the conflict feature vector and the four baseline feature vectors using a similarity algorithm. Compare the values ​​of the four vector similarities and determine the memory conflict type of the baseline feature vector corresponding to the maximum vector similarity as the memory conflict type corresponding to the current conflict feature vector. Simultaneously, use the maximum value among the four vector similarities as the confidence score output by the conflict type classifier. If the obtained confidence level is lower than the preset confidence level threshold, the current conflict feature vector is marked as manually reviewed, and the memory conflict type is not determined at this time. The memory conflict type will be manually confirmed and labeled later.

4. The multi-agent memory conflict arbitration method based on version vectors and policy libraries as described in claim 1, characterized in that: In step S3, the strategy matching and arbitration module matches the corresponding arbitration strategy from the pre-built strategy library based on the memory conflict type obtained in step S2 and performs the arbitration operation; the strategy library contains the association mapping relationship between memory conflict types and various arbitration strategies, and stores the success rate weight of various arbitration strategies corresponding to different memory conflict types; Based on the association mapping relationship, select the available arbitration strategies under the current memory conflict type, sort them according to the success rate weight of each arbitration strategy, select the arbitration strategy with the highest success rate weight as the final execution strategy, and execute the arbitration operation; The arbitration strategies include voting strategies, merging strategies, retaining multiple versions strategies, and marking uncertainties strategies; After each memory conflict arbitration is completed and a conflict resolution log is generated, the conflict resolution log is archived and stored as a historical arbitration case.

5. The multi-agent memory conflict arbitration method based on version vectors and policy libraries as described in claim 1, characterized in that: The specific details of step S4 are as follows: S41: Generate a list of memories; Once the conflict resolution log is generated and the arbitration-related process is completed, the feedback learning module is triggered; the memory of this conflict is extracted from the generated conflict resolution log and a memory list is generated. S42: Verify the arbitration effect and generate feedback learning samples; The trigger feedback learning module obtains the task execution results output by the multi-agent collaborative system after executing business tasks through the interactive interface. If the task execution results are normal and meet expectations, the arbitration operation corresponding to this memory conflict is marked as correct; otherwise, the corresponding arbitration operation is marked as pending verification. For arbitration operations marked as pending verification, manual review or feedback learning modules analyze the reasons for business task anomalies to determine whether they are caused by arbitration operations. If the decision is made due to arbitration, the feedback learning module will label the corresponding conflict feature vector, error handling result, and correct handling result together and add them to the training database as feedback learning samples. The error handling result is the arbitration result in the conflict handling log. The correct handling result is the correct memory version and correct memory conflict type corresponding to this memory conflict, obtained after manual review and re-deduction by the system. S43: Update the success rate weights of each arbitration strategy in the strategy library; The success rate of each arbitration strategy in the statistical strategy library under different memory conflict types; If the success rate of a certain arbitration strategy is lower than a preset threshold under a certain type of memory conflict, the success rate weight of that arbitration strategy will be reduced, and the weights of other arbitration strategies will be increased accordingly. Historical arbitration cases are analyzed in batches according to a preset cycle to generate conflict resolution effectiveness reports.

6. A multi-agent memory conflict arbitration system based on version vectors and policy libraries for implementing the method of any one of claims 1-5, characterized in that: It includes a memory version vector generation module, a conflict detection and classification module, a policy matching and arbitration module, and a feedback learning module; The memory version vector generation module is used to generate memory version vectors, establish associated indexes, and form memory version chains. The conflict detection and classification module is used to filter candidate memories to form a candidate conflict set, construct a conflict feature vector, and determine the type of memory conflict. The strategy matching and arbitration module selects an arbitration strategy based on the memory conflict type and the success rate weight of the strategy library, and generates a conflict handling log. The feedback learning module verifies the arbitration effect, collects error samples, and updates the policy weights.

Images