Intelligent dialogue system for time sequence memory management and task distribution

Abstract

The invention discloses an intelligent dialogue system for time sequence memory management and task distribution, which is based on CrewAI multi-agent collaboration, combines time sequence memory updating and multi-agent collaboration work, introduces an independent memory unit in order to solve the problem of low memory efficiency management, and uses an undirected graph structure to store learned information. Storing each piece of information as a node in the graph; meanwhile, memory is increased, modified and forgotten according to a time sequence, so that the problem that important information is forgotten is relieved to a certain extent; in order to solve the problem of insufficient task interaction, a multi-agent cooperation mechanism is introduced, user demand understanding, task execution and task verification are respectively submitted to three different agents for independent work, and through independent work and complete information exchange among the three agents, information synchronization and state sharing among modules are realized, information islands are reduced, and the task interaction efficiency is improved. And decision accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a multi-agent collaborative dialogue system in the field of artificial intelligence, and particularly to an intelligent dialogue system for temporal memory management and task distribution based on the multi-agent framework CrewAI. Background Art

[0002] Multi-agent dialogue is one of the important solutions for realizing complex task processing and long-range dialogue in the field of artificial intelligence. By handing over knowledge extraction and learning, task decomposition, task execution and task verification to different agents independently, and having an independent memory management system responsible for memory forgetting and updating, the current multi-agent dialogue system has the following problems: 1. Problem of low memory efficiency management: Low memory efficiency is a common problem in artificial intelligence dialogue systems. The memory management of existing dialogue systems not only needs to store dialogue content, but also needs to maintain multi-dimensional information such as context relationships, user states, and task progress. Compared with traditional information storage systems, it consumes a large amount of computing resources and storage space. At the same time, as the number of dialogue turns increases, the system needs to balance the integrity of information and processing efficiency under limited computing resources. A simple sliding window mechanism will cause the loss of important historical information, while full-scale storage will result in low retrieval efficiency and resource waste. In addition, when the dialogue scenario changes, the memory management strategy needs to be adjusted accordingly, which further increases the difficulty of system maintenance. Therefore, multi-agent dialogue will face problems such as heavy storage burden and difficulty in balancing resources and efficiency.

[0003] 2. Problem of insufficient task interaction: Existing dialogue systems often split functions such as user requirement understanding, memory management, and task execution. Although this modular design is convenient for management and maintenance in engineering implementation, it also brings serious problems of functional fragmentation. Each module is only connected through parameter sharing or simple message passing, and this shallow information interaction method cannot meet the requirements of complex dialogue scenarios. In multi-task collaborative processing, the challenges faced by the system are far greater than those in single-task processing. First, it is necessary to accurately identify and process the complex dependencies between tasks. For example, the output of some tasks may be the necessary input for other tasks. Second, resource competition occurs when multiple tasks are running simultaneously. The system needs to reasonably allocate limited resources such as computing resources and memory space, which increases the complexity of task scheduling. In dealing with complex multi-turn conversations, the information island effect between task modules becomes increasingly obvious. For example, the natural language understanding module may obtain some implicit intentions of the user, but due to the lack of an effective information sharing mechanism, these important information cannot be promptly transmitted to the task planning or execution module. Similarly, the experience and context information accumulated during the processing of the task execution module are also difficult to feedback to the requirement understanding module, resulting in the system being unable to perform global optimization and adjustment. This information island effect will ultimately lead to deviations in system decisions.

[0004] Problem status and solution: Updating the agent's memory according to the time sequence is a relatively mature method to solve the problem of memory forgetting in long-term conversations of agents. There are also relevant multi-agent collaborative work models to solve the problem of insufficient task interaction of agents, but there is no case in the industry that combines updating memory according to the time sequence and multi-agent collaborative work. Summary of the Invention

[0005] Aiming at the problems existing in the above-mentioned prior art, the object of the present invention is to combine updating memory according to the time sequence and multi-agent collaborative work. To solve the problem of low memory efficiency management, an independent memory unit is introduced, and an undirected graph structure is used to store the learned information, and each piece of information is stored as a node in the graph. At the same time, memory addition, modification, and forgetting are performed according to the time sequence, which alleviates problems such as forgetting of important information to a certain extent. To solve the problem of insufficient task interaction, a multi-agent collaborative mechanism is introduced. The user requirement understanding, task execution, and task verification are respectively handed over to three different agents for independent work. Through the independent work and complete information exchange between the three agents, information synchronization and status sharing between modules are realized, information islands are reduced, and decision-making accuracy is improved.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: The present invention is designed based on the multi-agent cooperation framework CrewAI, and includes three types of agents, namely AI user, AI assistant, and decider, and a memory unit. After the information input by the user is processed by the AI User, basic features such as user portraits, semantic information, and task requirements are obtained, and these basic feature information is stored in the sequential memory unit in the memory system for management; in the sequential memory unit, every time a round of tasks is completed, it will be updated according to the memory update formula, making up for problems such as memory redundancy and memory forgetting caused by traditional single storage; after the task is decomposed by the AI user, it is input into the task queue to solve the problems of task dependence and resource competition. Based on the multi-agent cooperation mechanism, the AI assistant is responsible for executing the task, and the execution result is passed to the decider for judging the correctness of the execution result. At the same time, the information learned from the execution task is passed to the memory unit for memory update. If the result is correct, the information is passed to the AIuser to continue executing the next task. If the result is incorrect, this task is executed again. The system is optimized through the loop mechanism of repeated execution of task learning - task distribution - task execution - task feedback, reducing system resource consumption while improving task processing performance, and repeating the process of task distribution, execution verification, and memory update until the long-term dialogue coherence goal is achieved.

[0007] This system includes a user interaction and information processing module, a memory unit, a task execution module, and a result verification module; The user interaction and information processing module: First, it processes the information input by the user through the AI User; second, it decomposes tasks and expands knowledge; third, it prepares for task distribution; The memory unit: After the AI user extracts and classifies the user information, a knowledge graph is constructed, and after each round of tasks is completed, it is updated according to the memory formula; The task execution module: After the task is decomposed by the AI user, it is input into the task queue, and then the AIassistant is responsible for executing the task, passing the execution result to the result verification module, and at the same time passing the information learned from the execution task to the memory unit for memory update; The result verification module: Judges the correctness of the execution result through the decider. If the result is correct, the information is passed to the AI user to continue executing the next task. If the result is incorrect, this task is executed again.

[0008] As a preferred mode of the present invention, in the user interaction and information processing module: The AI user is responsible for extracting information from the user side. The system will parse the user input to obtain basic feature information, store these feature information in the memory unit for management, and then perform task decomposition and knowledge expansion. The system decomposes complex tasks into executable subtasks, assigns priorities to each subtask. For these subtasks, the system performs knowledge expansion through the RAG technology, matches similar cases and solutions in the knowledge graph, adds the newly obtained knowledge as leaf nodes to the knowledge graph, and updates the memory and integrates with the existing knowledge together with the original knowledge graph; finally, it is to prepare for task distribution. The system will retrieve relevant information from the memory unit according to the current task context, extract the detailed information, execution requirements and knowledge support of the subtasks, and construct these complete task information packets into a priority task queue and pass them to the subtask queue.

[0009] As a preferred embodiment of the present invention, the construction of the knowledge graph includes: Extract the user portrait and task content from the user interaction and information processing module as two core nodes in the knowledge graph, and connect them with an edge to establish an initial relationship. Then, the AI user learns relevant knowledge through the RAG technology, and adds each newly obtained knowledge point as a leaf node to the graph. These leaf nodes are connected to the task content node; in the subsequent processing process, when the AI user retrieves information from the memory unit, search paths will be formed between different nodes. Whenever a node is searched to another node, a new edge is established between these two nodes, so that the graph can dynamically reflect the association between nodes; finally, when generating a specific task, the system will identify all the knowledge nodes used, and connect these nodes actually used in the task generation process in sequence through edges to form a complete sub-graph structure related to the task; at the same time, in the process of the AI assistant executing each round of tasks, the AI assistant will insert the learned knowledge as leaf nodes into the sub-graph related to this task in the knowledge graph, and the knowledge graph will be updated for the whole graph after each round.

[0010] The forgetting probability of a node is calculated by the following formula: , where S is the node importance and T is the time decay factor.

[0011] The node importance S is calculated by the following formula: , where C is the node centrality, and the calculation formula is: , Among them, DC is the degree centrality of the node, BC is the betweenness centrality of the node, and EC is the eigenvector centrality of the node. The more connections the node has with its adjacent nodes and the closer the relationship is in the graph, the larger the degree centrality value of the node, the larger the betweenness centrality value of the node, and the larger the eigenvector centrality value of the node. Thus, the node centrality value is larger.

[0012] Among them, F is the access frequency of the node in the last 3 rounds. R is the weight of the most recent access time, and the calculation formula is: , The closer the time when the node was accessed is to the current time, the smaller the difference between the current time and the last access time, and thus the larger the time weight. D is the calculation formula for the node dependence degree: , The more degrees the node has and the more active nodes there are around it, the greater the probability that the node will be accessed and the higher the node dependence degree. In the formula, The sum of is 1. By calculating the node importance, it is possible to screen out which nodes are more likely to be forgotten.

[0013] T is the time decay factor, and the calculation formula is: , Among them, k is the decay rate parameter, and m is the number of rounds that the node has not been accessed. The more rounds that the node has not been accessed, the larger the value of m and the smaller the value of T. Therefore, when a certain node is less important, the node importance of this node is lower, the time decay factor is higher, and the probability of being forgotten is greater. For core nodes, if the relevance to the current task is greater than 0.5, forgetting is temporarily prohibited. If the degree of the core node is greater than the average node degree, forgetting stops.

[0014] As a preferred embodiment of the present invention, in the task execution module: First, when the system receives an execution command, the AI assistant will extract key information from the subtask queue. This information includes not only specific subtask requirements and constraints, but also relevant knowledge support retrieved from the memory unit, including historical execution experience, solutions, and precautions, etc.; then, the AI assistant independently completes task processing. Finally, the AI assistant will pass the complete execution result to the decider for quality assessment and validity verification; at the same time, the AI assistant will pass the new knowledge and experience obtained during the task execution process (successful solutions, lessons learned from failures, and optimized strategies, etc.) to the memory unit for knowledge update.

[0015] As a preferred embodiment of the present invention, in the judgment module: The Decider is a pre-trained model used to determine whether the execution result of the AI assistant meets the task requirements. After receiving the execution result submitted by the AI assistant, the Decider will first extract the key features of the original task requirements and the execution result, and then, based on the preset evaluation criteria, conduct a multi-dimensional quality assessment of the execution result. If the execution result meets the requirements, the information will be passed to the AI user to enable it to generate the next task. If the execution result does not meet the requirements, the information will be passed to the AI user to enable it to regenerate the current task by combining the newly learned information.

[0016] The multi-dimensional quality assessment includes functional integrity, logical correctness, efficiency rationality, and the degree of matching with user requirements.

[0017] Compared with the prior art, the main advantages of the present invention are as follows: (1) The present invention adopts a multi-agent collaborative architecture, where user interaction, task execution, and result verification are respectively responsible by dedicated agents, and knowledge sharing is achieved through the memory unit constructed by the graph neural network. This design enables the system to accurately understand user requirements, efficiently execute tasks, and at the same time ensure the quality of the execution results. In particular, by processing user input through the RAG technology and updating the knowledge graph, the knowledge acquisition and utilization capabilities of the system are significantly improved; (2) The present invention introduces a task queue management mechanism. Through the AI User, task decomposition and distribution are carried out. Combining the historical experience in the memory unit, complex tasks can be decomposed into manageable subtasks. This design improves the efficiency and accuracy of task execution on the one hand, and on the other hand, ensures the quality of the execution results through the verification mechanism of the Decider, effectively solving the limitations of traditional dialogue systems in dealing with complex tasks; (3) The present invention introduces a memory management mechanism based on a graph structure. On the one hand, through node importance scoring and dynamic forgetting mechanism, the system can adaptively retain key information while timely clearing redundant data, reducing the resource consumption of memory storage. On the other hand, through the hierarchical structure and relationship modeling of the memory graph, effective organization and rapid retrieval of knowledge are realized, solving the problems of long-term dependence and difficult maintenance of knowledge association in traditional memory systems; (4) The present invention adopts a flexible verification feedback mechanism. The Decider evaluates the task execution result and decides whether to re-execute the task according to the evaluation result. This mechanism not only ensures the output quality, but also can continuously optimize the task execution strategy through the feedback loop, improving the overall performance of the system. At the same time, the new knowledge obtained during the execution process will be updated to the memory unit in a timely manner, realizing the continuous learning and evolution of the system.

[0018] (5) The present invention as a whole adopts a modular and extensible design concept. On the one hand, the responsibilities of each agent are clear and the interfaces are standardized, which facilitates model replacement and function expansion according to actual needs. On the other hand, through a unified communication protocol and memory sharing mechanism, efficient cooperation between agents is achieved, improving the overall performance and maintainability of the system. Description of the Drawings

[0019] Figure 1 It is a schematic structural diagram of the intelligent dialogue system in this embodiment. Specific Embodiments

[0020] To better understand the present invention, the technical solutions of the present invention will be further explained below in conjunction with the drawings in the specification and specific embodiments.

[0021] As Figure 1 shown, this embodiment proposes an intelligent dialogue system for time-series memory management and task distribution based on CrewAI's multi-agent cooperation. Through the collaborative work of multiple agents, this system constructs memory units in combination with the properties of the graph, achieving intelligent processing of complex tasks, continuous accumulation of knowledge, and effective control of task execution quality. The system uses GPT-3.5-turbo as the interaction and execution agent, Mistral-7B as the decision-making agent, conducts knowledge learning and expansion through the RAG technology, and realizes knowledge update through the properties of the graph to improve the overall performance and reliability of the system. The intelligent dialogue system for time-series memory management and task distribution based on CrewAI's multi-agent cooperation mainly includes a user interaction and information processing module, a memory unit, a task execution module, and a result verification module.

[0022] User Interaction and Information Processing Module: First step, the AI user is responsible for extracting information from the user side, and the system will parse the user input and classify the input information into categories such as user portraits, task requirements, and constraints. The second step is task decomposition and knowledge expansion. The system decomposes complex tasks into executable subtasks and assigns priorities to each subtask. For these subtasks, the system conducts knowledge expansion through the RAG technology, matches similar cases and solutions in the knowledge graph, adds the newly obtained knowledge as leaf nodes to the knowledge graph, and follows the original knowledge graph for memory update and integration with existing knowledge. The third step is task distribution preparation. The system will retrieve relevant information from the memory unit according to the current task context, extract the detailed information, execution requirements, and relevant knowledge support of the subtasks, and construct these complete task information packets into a priority task queue and pass them to the subtask queue.

[0023] Memory Unit: After the AI user extracts and classifies user information, a knowledge graph is constructed. The user portrait and task content are extracted from the user input as two core nodes in the knowledge graph, and an edge is used to connect them to establish an initial relationship. Then, the AI user learns relevant knowledge through the RAG technology, and each new knowledge point obtained is added to the graph as a leaf node. These leaf nodes are mainly connected to the task content node. In the subsequent processing, when the AI user retrieves information from the memory unit, search paths are formed between different nodes. Whenever a node is searched to another node, a new edge is established between these two nodes, enabling the graph to dynamically reflect the associations between nodes. Finally, when generating a specific task, the system identifies all the knowledge nodes used, and the nodes actually used in the task generation process are sequentially connected by edges to form a complete sub-graph structure related to the task. At the same time, during the execution of each round of tasks by the AI assistant, the AI assistant inserts the learned knowledge as leaf nodes into the sub-graph related to the current task in the knowledge graph. After each round, the entire knowledge graph is updated.

[0024] The forgetting probability of a node is calculated by the following formula: , where S is the node importance and T is the time decay factor.

[0025] The node importance is calculated by the following formula: , where C is the node centrality, and the calculation formula is: , where DC is the degree centrality of the node, BC is the betweenness centrality of the node, and EC is the eigenvector centrality of the node. The more connections and closer relationships there are between the node and its adjacent nodes in the graph, the larger the degree centrality value of the node, the larger the betweenness centrality value of the node, and the larger the eigenvector centrality value of the node, thus the larger the node centrality value.

[0026] where F is the access frequency of the node in the last 3 rounds. R is the weight of the most recent access time, and the calculation formula is: , The closer the time when the node was accessed is to the current time, the smaller the difference between the current time and the last access time, and thus the larger the time weight. D is the node dependence degree, and the calculation formula is: , The more degrees the node has and the more active nodes there are around it, the greater the probability that the node will be accessed and the higher the node dependence degree. In the formula, The sum is 1. By calculating the node importance, it is possible to filter out which nodes are more likely to be forgotten.

[0027] T is the time decay factor, and the calculation formula is: , where k is the decay rate parameter and m is the number of rounds that the node has not been visited. The more rounds the node has not been visited, the larger the value of m and the smaller the value of T. Therefore, when a node is less important, its node importance is lower, the time decay factor is higher, and the probability of being forgotten is greater. For core nodes, if the relevance to the current task is greater than 0.5, forgetting is temporarily prohibited. If the degree of the core node is greater than the average node degree, forgetting stops. By implementing memory management, the system can store information more efficiently and improve information utilization.

[0028] Task execution module. First, when the system receives an execution command, the AI assistant extracts key information from the subtask queue, which includes not only specific subtask requirements and constraints but also relevant knowledge support retrieved from the memory unit, such as historical execution experience, solutions, and precautions. Second, the AI assistant independently completes task processing. Third, the AI assistant passes the complete execution result to the decider for quality assessment and effectiveness verification. At the same time, the AI assistant passes the new knowledge and experience obtained during the task execution, such as successful solutions, lessons learned from failures, and optimized strategies, to the memory unit for knowledge update, thereby continuously enhancing the system's execution ability and adaptability.

[0029] Judgment module. The Decider is a pre-trained model used to determine whether the execution result of the AI assistant meets the task requirements. When receiving the execution result submitted by the AI assistant, the Decider first extracts the key features of the original task requirements and the execution result, and then, based on the preset evaluation criteria, conducts a multi-dimensional quality assessment of the execution result, including functional integrity, logical correctness, efficiency rationality, and the degree of matching with user requirements. If the execution result meets the requirements, the information is passed to the AI user to enable it to generate the next task. If the execution result does not meet the requirements, the information is passed to the AI user to enable it to regenerate the current task in combination with the newly learned information.

[0030] The above embodiments are only used to illustrate the technical concept and features of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. It is not intended to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent dialogue system for temporal memory management and task distribution, based on CrewAI’s multi-agent collaboration, characterized by: It includes user interaction and information processing module, memory unit, task execution module and result verification module; The user interaction and information processing module: first, processes the information input by the user through AI User; second, performs task decomposition and knowledge expansion; third, prepares for task distribution; The memory unit: after the AI ​​user extracts and classifies the user information, it builds a knowledge graph and updates the formula according to the memory after each round of tasks is completed; The task execution module: After being decomposed by the AI ​​user, the task is input into the task queue, and then the AI ​​assistant is responsible for executing the task, passing the execution result to the result verification module, and at the same time passing the information learned from the task execution to the memory unit for memory update; The result verification module: determines the correctness of the execution result through the decider. If the result is correct, the information is transmitted to the AI ​​user to continue to execute the next task. If the result is incorrect, the task is re-executed.

2. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: The processing of the information input by the user through the AI ​​User is specifically as follows: AI user is responsible for extracting information from the user side. The system will parse the user input, obtain basic feature information, and store this basic feature information into the memory unit for management. The basic feature information includes user portrait, task requirements, semantic information and constraints.

3. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: The task decomposition and knowledge expansion are specifically as follows: The system breaks down complex tasks into executable subtasks and assigns a priority to each subtask. For these subtasks, the system uses RAG technology to expand knowledge, matches similar cases and solutions in the knowledge graph, adds the newly acquired knowledge as leaf nodes to the knowledge graph, and updates memory and integrates existing knowledge with the original knowledge graph.

4. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: The task distribution preparation is specifically as follows: The system retrieves relevant information from the memory unit according to the current task context, extracts detailed information, execution requirements, and knowledge support of the subtask, and builds these complete task information packages into priority task queues and passes them to the subtask queues.

5. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: The construction of the knowledge graph includes: User portraits and task contents are extracted from the user interaction and information processing modules as two core nodes in the knowledge graph, and they are connected by an edge to establish an initial relationship. Then, the AI ​​user learns relevant knowledge through RAG technology, and adds each new knowledge point as a leaf node to the graph, and these leaf nodes are connected to the task content nodes. In the subsequent processing, when the AI ​​user retrieves information from the memory unit, a search path is formed between different nodes. Whenever a node is searched from another node, a new edge is established between the two nodes, so that the graph can dynamically reflect the relationship between the nodes. Finally, when generating a specific task, the system will identify all the knowledge nodes used, and connect these nodes actually used in the task generation process in sequence through edges to form a complete task-related subgraph structure. At the same time, in the process of the AI ​​assistant executing each round of tasks, the AI ​​assistant will insert the learned knowledge as a leaf node into the subgraph related to this task in the knowledge graph, and the entire knowledge graph will be updated after each round.

6. The intelligent dialogue system for temporal memory management and task distribution according to claim 5, characterized in that: The forgetting probability of a node is calculated by the following formula: ， Where S is the node importance and T is the time decay factor; The node importance S is calculated by the following formula: ， Among them, C is the node centrality, and the calculation formula is: ， Among them, DC is the degree centrality of the node, BC is the betweenness centrality of the node, and EC is the eigenvector centrality of the node. The more connections a node has with its adjacent nodes in the graph, and the closer the relationship is, the greater the degree centrality value of the node, the greater the betweenness centrality value of the node, the greater the eigenvector centrality value of the node, and thus the greater the node centrality value; Among them, F is the access frequency of the node in the last three rounds, R is the weight of the most recent access time, and the calculation formula is: ， The closer the time when the node is accessed is to the current time, the smaller the difference between the current time and the last access time is, and thus the greater the time weight is; D is the calculation formula for the node dependency: ， The more degrees the node has and the more active nodes are around it, the greater the probability that the node will be visited and the higher the node dependence. In the formula, The sum of is 1. By calculating the importance of nodes, we can filter out which nodes are more likely to be forgotten. T is the time attenuation factor, and the calculation formula is: ， Where k is the decay rate parameter, m is the number of rounds that the node has not been visited. The more rounds the node has not been visited, the larger the value of m and the smaller the value of T. Therefore, when a node is less important, the lower the node importance, the higher the time decay factor, and the greater the probability of being forgotten. For core nodes, if the relevance to the current task is greater than 0.5, forgetting is temporarily prohibited. If the degree of the core node is greater than the average node degree, forgetting is stopped.

7. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: In the task execution module: First, when the system receives an execution command, the AI ​​assistant will extract key information from the subtask queue, which includes not only specific subtask requirements and constraints, but also relevant knowledge support retrieved from the memory unit; then, the AI ​​assistant completes the task processing independently, and finally, the AI ​​assistant will pass the complete execution results to the decider for quality assessment and validity verification; at the same time, the AI ​​assistant will pass the new knowledge and experience gained during the task execution to the memory unit for knowledge updating.

8. The intelligent dialogue system for temporal memory management and task distribution according to claim 7, characterized in that: The relevant knowledge support includes historical execution experience, solutions and precautions; the new knowledge and experience include successful solutions, lessons from failure and optimized strategies.

9. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: In the judgment module: Decider is a pre-trained model used to determine whether the execution results of the AI ​​assistant meet the task requirements. After receiving the execution results submitted by the AI ​​assistant, Decider will first extract the key features of the original task requirements and the execution results, and then perform a multi-dimensional quality assessment of the execution results based on the preset evaluation criteria. If the execution results meet the requirements, the information will be passed to the AI ​​user to enable it to continue to generate the next task. If the execution results do not meet the requirements, the information will be passed to the AI ​​user to enable it to regenerate the task based on the newly learned information.

10. The intelligent dialogue system for temporal memory management and task distribution according to claim 7, characterized in that: The multi-dimensional quality assessment includes functional completeness, logical correctness, efficiency rationality and matching degree with user needs.

11. The intelligent dialogue system for temporal memory management and task distribution according to claim 1, characterized in that: The system works through collaboration among multiple intelligent agents, including AI User based on GPT-3.5-turbo, AI Assistant based on GPT-3.5-turbo, and Decider based on Mistral-7B.

Images