Long-term memory system and method based on memory-enhanced multi-agent collaborative mind model

Abstract

The application discloses a long-term memory system based on a memory-enhanced multi-agent collaborative mind model, and adopts a binary layered mixed storage architecture; the system comprises a cognitive model memory module CM, and the cognitive model memory module CM is connected with an execution experience memory module EE; the cognitive model memory module CM stores structured cognitive information of team tasks according to a three-layer structure of a task unit, a deliverable and an atomic activity; the execution experience memory module EE stores full-process operation data in a layered mode, and reversely optimizes the storage content of the cognitive model memory module CM. The application further discloses a running method of the long-term memory system based on the memory-enhanced multi-agent collaborative mind model. The application solves the problems of insufficient structured degree of long-term memory of an agent and low matching accuracy of task instructions in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence agent technology, specifically relating to a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, and also to a method for operating the long-term memory system based on a memory-enhanced multi-agent collaborative mental model. Background Technology

[0002] Currently, AI agents developed for team collaboration scenarios are mostly built on large language models and directed acyclic graphs, and are combined with retrieval enhancement technology to achieve memory retrieval. This technical model has many problems.

[0003] First, each process node relies on manual pre-definition, resulting in a static structure that struggles to respond to dynamic command changes and lacks the ability to accumulate historical experience and self-update. Second, the memory content invoked for retrieval enhancement generation is mostly unstructured text sets, leading to slow processing speeds and high computational costs when repeatedly reasoning about the same content. Third, the lack of explicit, structured team-shared mental models (a consensus cognitive schema for team collaboration) prevents AI from autonomously completing collaborative actions such as task decomposition, activity arrangement, and deliverable verification, resulting in significant errors in command comprehension and inconsistent accuracy of deliverables.

[0004] Existing AI long-term memory designs mostly focus on storing conversational information and personalized interactive content of a single agent, without designing dedicated memory structures, update rules, and decision-making linkage logic for team consensus cognition, thus failing to support long-term collaborative tasks involving multiple agents and multiple stakeholders. Summary of the Invention

[0005] The purpose of this invention is to provide a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, which solves the problems of insufficient structure of agent long-term memory and low accuracy of task instruction matching in the prior art.

[0006] Another objective of this invention is to provide a method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model.

[0007] The technical solution adopted in this invention is a long-term memory system based on a memory-enhanced multi-agent collaborative mental model. This system adopts a binary hierarchical hybrid storage architecture. The system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE). The cognitive model memory module (CM) stores structured cognitive information of team tasks according to a three-layer structure of task units, deliverables, and atomic activities. The execution experience memory module (EE) stores the entire process operation data in a hierarchical manner and reverse-optimizes the storage content of the cognitive model memory module (CM).

[0008] The invention is further characterized by: The cognitive model memory module (CM) includes a task unit storage unit, which stores information about the sub-tasks of the team's work after decomposition, the mapping relationship between task unit identifiers and agents; the sub-task information includes task identifiers, descriptions, status identifiers, task dependencies, and task prompts.

[0009] The cognitive model memory module (CM) also includes a deliverable storage unit, which stores the standard deliverable information corresponding to each task unit, the mapping relationship between the deliverable identifier and the corresponding task unit identifier, and the standard deliverable information, including the deliverable identifier, template, quality standard, delivery and acceptance entity, and non-conforming iteration rules.

[0010] The cognitive model memory module (CM) also includes atomic activity storage units, which are used to store the indivisible execution action information corresponding to each deliverable, the mapping relationship between activity identifiers and corresponding deliverable identifiers; the indivisible execution action information includes activity identifiers, input-output sets, execution rules, execution subjects, and agent operation parameters.

[0011] The Cognitive Model Memory Module (CM) also includes a team member storage unit, which stores information about the people and agents involved in the collaboration. This information includes member identification, type, role division, ability description, access permissions, and decision preferences.

[0012] The Execution Experience Memory Module (EE) includes a runtime record storage unit, which is used to retain the full-link execution log of each task and bind it to the corresponding cognitive model entity through an identifier. The full-link execution log includes session information, hit tasks, activity call sequence, time consumption, and failure reasons.

[0013] The Execution Experience Memory Module (EE) also includes a user feedback storage unit, which is used to collect and associate user evaluations, problem types, and specific opinions on each task execution, and automatically bind the corresponding running instance, hit task, and deliverable. The Execution Experience Memory Module (EE) also includes a Task Optimization Analysis Unit, which is used to periodically calculate task processing volume, hit rate, delivery accuracy, and average time consumption, identify high-frequency failed tasks, redundant activities, and unreasonable processes, and generate optimization proposals.

[0014] The Cognitive Model Memory Module (CM) adopts a hybrid storage architecture combining relational databases and vector databases; structured metadata is stored in the relational database; semantic embedding vectors of task descriptions, deliverable names, and activity rules are stored in the vector database; the relational database and the vector database are linked one-to-one by a globally unique identifier. The Execution Experience Memory (EE) module adopts a relational database combined with an object storage architecture; structured execution logs and deliverables are stored in the relational database, while unstructured data of file-type deliverables is stored in object storage; unstructured data is bound to the execution log through task round identifiers.

[0015] Another technical solution adopted in this invention is a method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, comprising the following steps: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the task to hit; S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality; S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit; S4. The execution log and user feedback storage unit collects opinions, and the task optimization and analysis unit compiles statistics and triggers the cognitive model memory module (CM) to update.

[0016] Another feature of the technical solution adopted in this invention is that: S1 specifically refers to: S1.1 The main collaborative agent receives natural language instructions sent by the user, completes the task intent and core keyword extraction operation, and filters out noise information. S1.2. Call the vector database to perform semantic similarity retrieval, recall the Top 10 candidate task units, and perform weighted scoring and sorting according to semantic similarity, recent call frequency, critical path priority and member decision preferences in the team member storage unit; S1.3 When the confidence level of a candidate task is not lower than 0.85, the corresponding task is triggered directly; when the confidence level of a candidate task is lower than 0.6, a request for instruction clarification is sent to the user; when the confidence level is between 0.6 and 0.85, all candidate tasks are displayed for the user to select and confirm. S2 specifically refers to: S2.1. Locate the unique task unit and its corresponding deliverable, create a task execution round, and record the data of the entire execution process; S2.2 Retrieve all atomic activities corresponding to the deliverable from the atomic activity storage unit, generate an executable activity sequence based on predefined dependencies, and make adaptive fine-tuning of the activity order based on real-time resource usage without violating dependency requirements; S2.3. Execute each atomic activity sequentially according to the activity sequence, and verify the output quality of each step; if the output quality does not meet the requirements, the rework process is automatically triggered; if the requirements still cannot be met after repeated processing, the manual intervention mechanism is triggered. S3 specifically refers to: S3.1 Summarize all output results of each atomic activity, match the template and quality standards in the deliverable storage unit, generate the final deliverable and return it to the user; S3.2 After the task work round is completed and closed, the execution log, deliverable files and user feedback data of this task are synchronously entered into the running record storage unit and user feedback storage unit in the execution experience memory module. S4 specifically refers to: S4.1 The task optimization analysis unit statistically analyzes operational indicators according to a preset cycle, and identifies high-frequency failure nodes, redundant activities, and unreasonable dependencies. S4.2. Memory update is triggered when any of the following conditions are met: core metrics are below the convergence benchmark for three consecutive days; deliverables fail acceptance or atomic activities time out; or users explicitly request corrections. S4.3 Generate targeted optimization proposals, which are then manually reviewed and tested with 10% traffic. If no performance degradation is detected, the update is officially released, while retaining historical versions and supporting rollback.

[0017] The beneficial effects of this invention are: The long-term memory system and method based on a memory-enhanced multi-agent collaborative mental model provided by this invention, compared with traditional retrieval enhancement schemes, can improve task hit rate by 11%, reduce instruction comprehension bias, and improve the accuracy of task instruction matching; delivery accuracy can be improved by more than 60%, reducing rework and manual intervention. After stable operation, the manual intervention rate can be controlled within 3%, improving the quality of deliverables; average delivery efficiency is improved by 4.4%, reducing the computational power consumption caused by repetitive reasoning and shortening task delivery time; based on execution experience data, task rules and activity orchestration can be continuously optimized to adapt to constantly changing business needs and support dynamic iteration of the mental model. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the design of the long-term memory system based on the memory-enhanced multi-agent collaborative mental model of the present invention; Figure 2 This is a design diagram of the hierarchical storage architecture of the present invention; Figure 3 This is a flowchart illustrating the memory retrieval and task execution process of this invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0020] The present invention provides a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1 As shown, Figure 1 As shown, the system adopts a binary hierarchical hybrid storage architecture, referencing the binary characteristics of human memory to distinguish between cognitive information used for task execution and experiential data used for iteration, balancing execution stability and evolutionary capability. The system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE). The CM stores structured cognitive information about team tasks according to a three-tiered structure of task units, deliverables, and atomic activities. Referring to the outcome-oriented principle of the Work Breakdown Structure (WBS), it transforms abstract team tasks into machine-parseable structured rules, achieving automatic conversion from task instructions to execution actions. The execution experience memory module (EE) stores the entire process data hierarchically, reverse-optimizing the storage content of the cognitive model memory module (CM) and possessing a closed-loop update mechanism. Execution experience can reverse-optimize the cognitive model, enabling continuous evolution of the mental model. Figure 2As shown, the Cognitive Model Memory Module (CM) includes a task unit storage unit, which stores information about the sub-tasks of the team's work after decomposition, as well as the mapping relationship between task unit identifiers and intelligent agents. Sub-task information includes task identifiers, descriptions, status identifiers, task dependencies, and task prompts. The CM also includes a deliverable storage unit, which stores information about the standard deliverables for each task unit, as well as the mapping relationship between deliverable identifiers and corresponding task unit identifiers. Standard deliverable information includes deliverable identifiers, templates, quality standards, delivery and acceptance entities, and rules for non-compliance iterations. The CM also includes atomic activities. The storage unit, specifically the atomic activity storage unit, stores the indivisible execution action information corresponding to each deliverable, as well as the mapping relationship between activity identifiers and corresponding deliverable identifiers. The indivisible execution action information includes activity identifiers, input / output sets, execution rules, and execution subject and agent operation parameters. The cognitive model memory module (CM) also includes a team member storage unit, which stores information about the people and agents participating in the collaboration. This information includes member identifiers, types, roles, ability descriptions, access permissions, and decision preferences. The execution experience memory module (EE) includes a runtime record storage unit, which retains the entire execution process for each task. The execution log is bound to the corresponding cognitive model entity through an identifier; the end-to-end execution log includes session information, hit tasks, activity call sequence, time consumption, and failure reasons; the execution experience memory module (EE) also includes a user feedback storage unit, which collects and associates user evaluations, problem types, and specific opinions on each task execution, automatically binding the corresponding running instance, hit task, and deliverables; the execution experience memory module (EE) also includes a task optimization analysis unit, which periodically calculates task processing volume, hit rate, delivery accuracy, and average time consumption, identifies high-frequency failed tasks, redundant activities, and unreasonable processes, and generates optimization proposals; cognitive model memory... The Cognitive Model Memory (CM) module employs a hybrid storage architecture combining relational and vector databases. This leverages the advantages of both to simultaneously meet the needs of accurate semantic recall and structured data management. Structured metadata is stored in the relational database; semantic embeddings of task descriptions, deliverable names, and activity rules are stored in the vector database. The relational and vector databases are linked by globally unique identifiers. The Execution Experience Memory (EE) module uses a relational database combined with object storage. Structured execution logs and deliverables are stored in the relational database, while unstructured data (such as file-type deliverables) is stored in object storage. Unstructured data is bound to execution logs via task round identifiers. The structured storage of the Cognitive Model Memory (CM) module can be replaced with a graph database to adapt to task scenarios with complex dependencies. The weighted dimensions for semantic recall can be adjusted based on business scenarios, or new ranking dimensions can be added to suit different scenarios.The triggering conditions and cycle of memory updates can be flexibly configured according to business stability requirements.

[0021] The long-term memory system based on a memory-enhanced multi-agent collaborative mental model provided by this invention is designed based on the following principles: First, referencing the duality of human memory, cognition, and experience, we divide memory into two modules: cognition and experience, which respectively undertake the functions of task execution and iterative optimization. Second, referencing the outcome-oriented principle of work breakdown structure, the team's tasks are broken down from top to bottom into three layers: task units, deliverables, and atomic activities, with results anchored at each layer to ensure clear execution objectives. Third, taking into account the operational characteristics of AI agents, a hybrid storage architecture is adopted to balance semantic recall capabilities and structured management capabilities, and a closed-loop update mechanism is used to achieve dynamic optimization of the mental model.

[0022] The performance comparison test results of the long-term memory system based on a memory-enhanced multi-agent collaborative mental model provided by this invention are as follows: In an enterprise data quality management scenario, 200 business instructions were selected for comparative testing, using the traditional LangChain combined with RAG solution as the baseline. The test results are as follows: 1. Task hit rate: The baseline solution was 88%, and after applying this system, it was 99%, an improvement of 11 percentage points; 2. Delivery accuracy: The baseline accuracy was 38%, while after applying this system, it was 98%, an improvement of 60 percentage points; 3. Average delivery time: The baseline solution had an average delivery time of approximately 8.86 seconds per task, while the solution using this system had an average delivery time of approximately 8.47 seconds, a reduction of 4.4%. 4. Manual intervention rate: The baseline scheme has no dynamic optimization capability. After the system was officially launched, the manual intervention rate was ≤3% for two consecutive weeks.

[0023] The present invention provides a method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, comprising the following steps: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the task to hit; S1 specifically refers to: S1.1 The main collaborative agent receives natural language instructions sent by the user, completes the task intent and core keyword extraction operation, and filters out noise information. S1.2. Call the vector database to perform semantic similarity retrieval, recall the Top 10 candidate task units, and perform weighted scoring and sorting according to semantic similarity, recent call frequency, critical path priority and member decision preferences in the team member storage unit; S1.3 When the confidence level of a candidate task is not lower than 0.85, the corresponding task is triggered directly; when the confidence level of a candidate task is lower than 0.6, a request for instruction clarification is sent to the user; when the confidence level is between 0.6 and 0.85, all candidate tasks are displayed for the user to select and confirm. S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality; S2 specifically refers to: S2.1. Locate the unique task unit and its corresponding deliverable, create a task execution round, and record the data of the entire execution process; S2.2 Retrieve all atomic activities corresponding to the deliverable from the atomic activity storage unit, generate an executable activity sequence based on predefined dependencies, and make adaptive fine-tuning of the activity order based on real-time resource usage without violating dependency requirements; S2.3. Execute each atomic activity sequentially according to the activity sequence, and verify the output quality of each step; if the output quality does not meet the requirements, the rework process is automatically triggered; if the requirements still cannot be met after repeated processing, the manual intervention mechanism is triggered. S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit; S3 specifically refers to: S3.1 Summarize all output results of each atomic activity, match the template and quality standards in the deliverable storage unit, generate the final deliverable and return it to the user; S3.2 After the task work round is completed and closed, the execution log, deliverable files and user feedback data of this task are synchronously entered into the running record storage unit and user feedback storage unit in the execution experience memory module. S4. The execution log is stored in the running record storage unit and the user feedback storage unit collects opinions. The task optimization analysis unit periodically collects statistical indicators and triggers the cognitive model memory module (CM) to update. S4 specifically refers to: S4.1 The task optimization analysis unit statistically analyzes operational indicators according to a preset cycle, and identifies high-frequency failure nodes, redundant activities, and unreasonable dependencies. S4.2. Memory update is triggered when any of the following conditions are met: core metrics are below the convergence benchmark for three consecutive days; deliverables fail acceptance or atomic activities time out; or users explicitly request corrections. S4.3 Generate targeted optimization proposals, which are then manually reviewed and tested with 10% traffic. If no performance degradation is detected, the update is officially released, while retaining historical versions and supporting rollback.

[0024] Example 1 The long-term memory system proposed in this embodiment is based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1 As shown, the system adopts a binary hierarchical hybrid storage architecture. The system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE). The cognitive model memory module (CM) stores the structured cognitive information of team tasks in a three-layer structure of task units, deliverables, and atomic activities. The execution experience memory module (EE) stores the full-process operation data in a hierarchical manner and reverse-optimizes the storage content of the cognitive model memory module (CM).

[0025] Example 2 The long-term memory system proposed in this embodiment is based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1 As shown, the system adopts a binary hierarchical hybrid storage architecture; the system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE); the cognitive model memory module (CM) stores structured cognitive information of team tasks according to a three-layer structure of task units, deliverables, and atomic activities; the execution experience memory module (EE) stores the entire process operation data hierarchically, and reverse-optimizes the storage content of the cognitive model memory module (CM). Figure 2 As shown, the Cognitive Model Memory Module (CM) includes a task unit storage unit, which stores information about the sub-tasks of the team's work after decomposition, the mapping relationship between task unit identifiers and intelligent agents, and the sub-task information including task identifier, description, status identifier, task dependencies, and task prompts. The CM also includes a deliverable storage unit, which stores information about the standard deliverables for each task unit, the mapping relationship between deliverable identifiers and corresponding task unit identifiers, and the standard deliverable information including deliverable identifiers, templates, quality standards, delivery and acceptance entities, and non-conforming iteration rules. Furthermore, the CM includes an atomic activity storage unit, which stores information about the indivisible execution actions for each deliverable, the mapping relationship between activity identifiers and corresponding deliverable identifiers, and the indivisible execution action information including activity identifiers, input / output sets, execution rules, execution entities, and intelligent agent operating parameters. Finally, the CM includes a team member storage unit, which stores information about the people and intelligent agents participating in the collaboration, including member identifiers, types, roles, ability descriptions, access permissions, and decision preferences.

[0026] Example 3 The long-term memory system proposed in this embodiment is based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1 As shown, the system adopts a binary hierarchical hybrid storage architecture; the system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE); the cognitive model memory module (CM) stores structured cognitive information of team tasks according to a three-layer structure of task units, deliverables, and atomic activities; the execution experience memory module (EE) stores the entire process operation data hierarchically, and reverse-optimizes the storage content of the cognitive model memory module (CM). Figure 2 As shown, the Cognitive Model Memory Module (CM) includes a task unit storage unit, which stores information about the sub-tasks of the team's work after decomposition, the mapping relationship between task unit identifiers and intelligent agents, and the sub-task information including task identifier, description, status identifier, task dependencies, and task prompts. The CM also includes a deliverable storage unit, which stores information about the standard deliverables for each task unit, the mapping relationship between deliverable identifiers and corresponding task unit identifiers, and the standard deliverable information including deliverable identifiers, templates, quality standards, delivery and acceptance entities, and rules for iteration in case of non-compliance. Furthermore, the CM includes an atomic activity storage unit, which stores information about the indivisible execution actions for each deliverable, the mapping relationship between activity identifiers and corresponding deliverable identifiers, and the indivisible execution action information including activity identifiers, input / output sets, execution rules, execution entities, and intelligent agent operating parameters. The CM also includes a team... The team member storage unit stores information about the people and agents participating in the collaboration. This information includes member identifiers, types, roles, capability descriptions, access permissions, and decision preferences. The Execution Experience Memory Module (EE) includes a runtime record storage unit, which retains the full-link execution log for each task and binds it to the corresponding cognitive model entity through identifiers. The full-link execution log includes session information, hit tasks, activity call sequence, time consumption, and failure reasons. The Execution Experience Memory Module (EE) also includes a user feedback storage unit, which collects and associates user evaluations, problem types, and specific opinions on each task execution, automatically binding them to the corresponding runtime instance, hit task, and deliverables. The Execution Experience Memory Module (EE) also includes a task optimization analysis unit, which periodically calculates task processing volume, hit rate, delivery accuracy, and average time consumption metrics, identifies high-frequency failed tasks, redundant activities, and unreasonable processes, and generates optimization proposals.

[0027] Example 4 The long-term memory system proposed in this embodiment is based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1As shown, the system adopts a binary hierarchical hybrid storage architecture; the system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE); the cognitive model memory module (CM) stores structured cognitive information of team tasks according to a three-layer structure of task units, deliverables, and atomic activities; the execution experience memory module (EE) stores the entire process operation data hierarchically, and reverse-optimizes the storage content of the cognitive model memory module (CM). Figure 2As shown, the Cognitive Model Memory Module (CM) includes a task unit storage unit, which stores information about the sub-tasks of the team's work after decomposition, the mapping relationship between task unit identifiers and intelligent agents, and the sub-task information including task identifier, description, status identifier, task dependencies, and task prompts. The CM also includes a deliverable storage unit, which stores information about the standard deliverables for each task unit, the mapping relationship between deliverable identifiers and corresponding task unit identifiers, and the standard deliverable information including deliverable identifiers, templates, quality standards, delivery and acceptance entities, and non-conforming iteration rules. Furthermore, the CM includes an atomic activity storage unit, which stores information about the indivisible execution actions for each deliverable, the mapping relationship between activity identifiers and corresponding deliverable identifiers, and the indivisible execution action information including activity identifiers, input / output sets, execution rules, execution entities, and intelligent agent operating parameters. Finally, the CM includes a team member storage unit, which stores information about the people and intelligent agents participating in the collaboration, including member identifiers, types, roles, ability descriptions, access permissions, and decision preferences. The Execution Experience Memory Module (EE) includes a runtime record storage unit. The line log storage unit is used to retain the full-link execution log of each task and bind it to the corresponding cognitive model entity through an identifier. The full-link execution log includes session information, hit tasks, activity call sequence, time consumption, and failure reasons. The Execution Experience Memory Module (EE) also includes a user feedback storage unit, which is used to collect and associate user evaluations, problem types, and specific opinions on each task execution, and automatically bind them to the corresponding running instance, hit task, and deliverables. The Execution Experience Memory Module (EE) also includes a task optimization analysis unit, which is used to periodically calculate task processing volume, hit rate, delivery accuracy, and average time consumption indicators, and identify high-risk metrics. For frequently failed tasks, redundant activities, and unreasonable processes, optimization proposals are generated. The Cognitive Model Memory Module (CM) adopts a hybrid storage architecture combining relational databases and vector databases. Structured metadata is stored in the relational database; semantic embeddings of task descriptions, deliverable names, and activity rules are stored in the vector database. The relational database and the vector database are linked one-to-one by a globally unique identifier. The Execution Experience Memory Module (EE) adopts a relational database combined with an object storage architecture. Structured execution logs and deliverables are stored in the relational database, while unstructured data for file-type deliverables is stored in the object storage. Unstructured data is bound to the execution logs through task round identifiers.

[0028] Example 5 The operating method of the long-term memory system based on the memory-enhanced multi-agent collaborative mental model proposed in this embodiment includes the following steps, according to the aforementioned long-term memory system based on the memory-enhanced multi-agent collaborative mental model: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the task to hit; S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality; S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit; S4. The execution log is stored in the running record storage unit, and the user feedback storage unit collects opinions. The task optimization analysis unit periodically collects statistical indicators and triggers the cognitive model memory module (CM) to update.

[0029] Example 6 The operating method of the long-term memory system based on the memory-enhanced multi-agent collaborative mental model proposed in this embodiment, according to the above-mentioned long-term memory system based on the memory-enhanced multi-agent collaborative mental model, such as... Figure 3 As shown, it includes the following steps: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the task to hit; S1 specifically refers to: S1.1 The main collaborative agent receives natural language instructions sent by the user, completes the task intent and core keyword extraction operation, and filters out noise information. S1.2. Call the vector database to perform semantic similarity retrieval, recall the Top 10 candidate task units, and perform weighted scoring and sorting according to semantic similarity, recent call frequency, critical path priority and member decision preferences in the team member storage unit; S1.3 When the confidence level of a candidate task is not lower than 0.85, the corresponding task is triggered directly; when the confidence level of a candidate task is lower than 0.6, a request for instruction clarification is sent to the user; when the confidence level is between 0.6 and 0.85, all candidate tasks are displayed for the user to select and confirm. S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality; S2 specifically refers to: S2.1. Locate the unique task unit and its corresponding deliverable, create a task execution round, and record the data of the entire execution process; S2.2 Retrieve all atomic activities corresponding to the deliverable from the atomic activity storage unit, generate an executable activity sequence based on predefined dependencies, and make adaptive fine-tuning of the activity order based on real-time resource usage without violating dependency requirements; S2.3. Execute each atomic activity sequentially according to the activity sequence, and verify the output quality of each step; if the output quality does not meet the requirements, the rework process is automatically triggered; if the requirements still cannot be met after repeated processing, the manual intervention mechanism is triggered. S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit; S3 specifically refers to: S3.1 Summarize all output results of each atomic activity, match the template and quality standards in the deliverable storage unit, generate the final deliverable and return it to the user; S3.2 After the task work round is completed and closed, the execution log, deliverable files and user feedback data of this task are synchronously entered into the running record storage unit and user feedback storage unit in the execution experience memory module. S4. The execution log is stored in the running record storage unit and the user feedback storage unit collects opinions. The task optimization analysis unit periodically collects statistical indicators and triggers the cognitive model memory module (CM) to update. S4 specifically refers to: S4.1 The task optimization analysis unit statistically analyzes operational indicators according to a preset cycle, and identifies high-frequency failure nodes, redundant activities, and unreasonable dependencies. S4.2. Memory update is triggered when any of the following conditions are met: core metrics are below the convergence benchmark for three consecutive days; deliverables fail acceptance or atomic activities time out; or users explicitly request corrections. S4.3 Generate targeted optimization proposals, which are then manually reviewed and tested with 10% traffic. If no performance degradation is detected, the update is officially released, while retaining historical versions and supporting rollback.

[0030] Example 7 This embodiment provides a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, such as... Figure 1 As shown, the system adopts a binary hierarchical hybrid storage architecture, which is divided into two parts: the cognitive model memory module (CM) and the execution experience memory module (EE). The CM and EE are interconnected. This architecture, referencing the binary nature of human memory, distinguishes between cognitive information used for task execution and experiential data used for iteration, balancing execution stability and evolutionary capability.

[0031] The Cognitive Model Memory Module (CM) stores structured cognitive information about team tasks in a three-tiered structure of task units, deliverables, and atomic activities. Referring to the outcome-oriented principles of the Work Breakdown Structure (WBS), it transforms abstract team tasks into machine-parseable structured rules, enabling automatic conversion from task instructions to execution actions. The Execution Experience Memory Module (EE) stores the entire process data in a hierarchical manner and can reverse-engineer the storage content of the Cognitive Model Memory Module (CM), possessing a closed-loop update mechanism to achieve continuous evolution of the mental model.

[0032] The Cognitive Model Memory Module (CM) includes four types of storage units: task unit storage unit, deliverable storage unit, atomic activity storage unit, and team member storage unit.

[0033] The task unit storage unit stores information about the sub-tasks of the team's work after decomposition, as well as the mapping relationship between task unit identifiers and intelligent agents. The sub-task information includes a task identifier, task description, status identifier, task dependencies, and task prompts. In the data governance scenario, task units include five categories: periodic data quality inspection, data quality report generation, problem data query, problem root cause analysis, and governance operation Q&A. Each task unit has a unique task identifier and records the dependencies between tasks. Data quality report generation depends on the completion of periodic data quality inspections.

[0034] The deliverable storage unit stores the standard deliverable information for each task unit, as well as the mapping relationship between deliverable identifiers and corresponding task unit identifiers. The standard deliverable information includes deliverable identifiers, templates, quality standards, delivery and acceptance entities, and non-compliance iteration rules. The deliverable for periodic data quality inspection tasks is problematic data tags. The template specifies the field format and output structure of the tagged data. The quality standard requires a tagging accuracy of no less than 98%, and the acceptance entity is the data governance manager.

[0035] The atomic activity storage unit stores the indivisible execution action information corresponding to each deliverable, as well as the mapping relationship between activity identifiers and corresponding deliverable identifiers. The indivisible execution action information includes the activity identifier, input / output set, execution rules, execution subject, and agent operation parameters. Each problem data labeling deliverable corresponds to a set of atomic activities, including a data query activity that reads data to be detected from a specified data source, a rule verification activity that checks data item by item according to preset quality rules, a semantic analysis activity that categorizes the causes of abnormal data, a template matching activity that organizes output according to a template format, and a report generation activity that generates the final labeling results. Each atomic activity has clearly defined inputs and outputs.

[0036] The team member storage unit is used to store information about the people and agents participating in the collaboration. This information includes member identifiers, types, roles, capability descriptions, access permissions, and decision preferences. For example, the data query agent's member identifier is AGENT_001, its type is a data query agent, its role is data reading and preprocessing, its capability description is support for SQL database queries and data cleaning, its access permissions are access to all data sources on the data governance platform, and its decision preference is to prioritize using cached data to improve query efficiency.

[0037] The Execution Experience Memory Module (EE) includes three types of storage units: execution record storage unit, user feedback storage unit, and task optimization analysis unit.

[0038] The execution log storage unit is used to retain the full-link execution log of each task, and is bound to the corresponding cognitive model entity through an identifier. The full-link execution log includes session information, hit task, activity call sequence, time consumption, and failure reason. The execution log storage unit stores the log record of a single task execution, including session ID, hit task identifier, activity call sequence, time consumption of each activity, and failure reason.

[0039] The user feedback storage unit is used to collect and associate user evaluations, problem types, and specific comments for each task execution, automatically binding them to the corresponding running instance, hit task, and deliverable. For example, if a user marks problematic data in a delivery as "Some abnormal data in the marking results were not detected," the problem type is "Missing Detection," and the specific comment is "It is recommended to add validation rules for null fields," this feedback is automatically bound to the corresponding running instance and deliverable.

[0040] The task optimization analysis unit is used to periodically analyze metrics such as task processing volume, hit rate, delivery accuracy, and average time consumption, identify frequently failed tasks, redundant activities, and unreasonable processes, and generate optimization proposals. Weekly statistics revealed that the rule validation activity had an average time consumption exceeding 10 seconds when the data volume exceeded 1 million records, identifying it as a performance bottleneck node and generating an optimization proposal: "It is recommended to add a batch processing mechanism to the rule validation activity."

[0041] like Figure 2 As shown, the Cognitive Model Memory (CM) module adopts a hybrid storage architecture combining relational databases and vector databases, simultaneously meeting the needs of accurate semantic retrieval and structured data management. Structured metadata is stored in the relational database; semantic embedding vectors of task descriptions, deliverable names, and activity rules are stored in the vector database; the relational database and the vector database are linked one-to-one by a globally unique identifier.

[0042] The following tables are created in the relational database of the Cognitive Model Memory Module (CM): `task_unit` table, with fields including task identifier (primary key), task name, task description, status identifier, task dependencies stored in JSON format, and task prompt template; `deliverable` table, with fields including deliverable identifier (primary key), corresponding task unit identifier (foreign key), deliverable template stored in JSON format, quality standard stored in JSON format, acceptance subject, and non-conforming iteration rules; `atomic_activity` table, with fields including activity identifier (primary key), corresponding deliverable identifier (foreign key), input set definition stored in JSON format, output set definition stored in JSON format, execution rules, execution subject, and agent running parameters stored in JSON format; and `team_member` table, with fields including member identifier (primary key), member type, role assignment, ability description, access permissions stored in JSON format, and decision preferences stored in JSON format. In the vector database, semantic embedding vectors are generated for the task description field of the task_unit table, the deliverable name field of the deliverable table, and the execution rule field of the atomic_activity table. The text-embedding-3-small model is used, with a vector dimension of 1536, and the vectors are stored in the corresponding collections. Each vector record is associated with the corresponding record in the relational database through a globally unique identifier.

[0043] The Execution Experience Memory (EE) module employs a relational database combined with an object storage architecture. Structured execution logs and deliverables are stored in the relational database, while unstructured data, such as file-type deliverables, are stored in object storage. Unstructured data is bound to the execution log through task round identifiers. The relational database of the EE module creates the `execution_log` table, with fields including: log identifier (primary key), session identifier, hit task identifier, execution start time, execution end time, activity call sequence stored as a JSON array, activity duration stored as a JSON object, failure reason, and execution status; the `user_feedback` table, with fields including: feedback identifier (primary key), corresponding execution log identifier (foreign key), hit task identifier, corresponding deliverable identifier, evaluation level, problem type, specific comments, and creation time; and the `optimization_proposal` table, with fields including: proposal identifier (primary key), analysis period, indicator type, current value, baseline value, problem description, suggested solution, status, and version number. In the object storage MinIO, file-type deliverables are organized and stored by Bucket. Bucket names are categorized by task type, and the object key format for each object is {task identifier} / {deliverable identifier} / {execution time}_{filename}. Files in object storage are associated with execution logs through task round identifiers, i.e., log identifiers in the execution_log table. When it is necessary to trace the complete execution of a task, the execution logs and user feedback are queried in the relational database using the log identifiers, and the corresponding deliverable file is located in object storage using the log identifiers, achieving complete traceability of end-to-end data.

[0044] The structured storage of the cognitive model memory module (CM) can be replaced with a graph database to adapt to task scenarios with complex dependencies. The weighted dimensions of semantic recall can be adjusted according to business scenarios, or new ranking dimensions can be added to suit the scenarios. The triggering conditions and cycle of memory updates can be flexibly configured according to business stability requirements.

[0045] Example 8 This embodiment provides a method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, and is based on the long-term memory system based on a memory-enhanced multi-agent collaborative mental model in Embodiment 7, such as... Figure 3 As shown, it includes the following steps: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the tasks.

[0046] S1 specifically refers to: S1.1 The main collaborative agent receives natural language commands from the user, extracts the task intent and core keywords, and filters out noise. For example, if the user inputs the command "Check the data quality for this month," the main collaborative agent extracts the task intent as generating a data quality report, the core keywords as "this month" and "data quality," and filters out filler words and other noise.

[0047] S1.2. Perform semantic similarity retrieval using a vector database to retrieve the Top 10 candidate task units. Convert the preprocessed instruction text into vectors using the semantic embedding model text-embedding-ada-002. Calculate the cosine similarity between the vector text and the description vectors of each task unit stored in the vector database. Sort the candidate task units from highest to lowest similarity and select the top 10. Then, weight the scores and rank the candidates based on semantic similarity, recent call frequency, critical path priority, and member decision preferences recorded in the team member storage units. The weighting formula is: Overall Score = 0.5 × Semantic Similarity + 0.2 × Normalized Recent Call Frequency + 0.2 × Normalized Critical Path Priority + 0.1 × Member Decision Preference Matching Degree. Where, recent call frequency counts the number of times each task has been triggered in the past 7 days; critical path priority assigns higher weights to tasks on the critical path in the task dependency graph; and member decision preferences are calculated based on the preference information recorded in the team member storage units.

[0048] S1.3 When the confidence level of a candidate task is not lower than 0.85, the corresponding task is triggered directly; when the confidence level of a candidate task is lower than 0.6, a request for instruction clarification is sent to the user; when the confidence level is between 0.6 and 0.85, all candidate tasks are displayed for the user to select and confirm.

[0049] S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality.

[0050] S2 specifically refers to: S2.1. Lock the unique task unit and its corresponding deliverable, create a task execution round, and record the data throughout the entire execution process. Lock the task unit "Data Quality Report Generation" and its corresponding deliverable "Data Quality Inspection Report", create an execution round ID, and initialize the execution log recorder.

[0051] S2.2 Retrieve all atomic activities corresponding to the deliverable from the atomic activity storage unit and generate an executable activity sequence based on predefined dependencies. The atomic activities corresponding to the data quality inspection report include data query, rule validation, semantic analysis, template matching, and report generation. The predefined dependencies are: data query before rule validation, rule validation before semantic analysis, semantic analysis before template matching, and template matching before report generation. Simultaneously, considering real-time resource usage, the activity order is adaptively fine-tuned without violating dependency requirements. When the rule validation activity has high computational resource usage while the semantic analysis activity has idle resources and there is no data dependency between the two, the semantic analysis activity is executed first.

[0052] S2.3. Execute each atomic activity sequentially according to the activity sequence, and verify the output quality of each step. Execute the data query activity to read all business data for the month from the data source and verify that the output dataset is not empty and that all fields are complete; execute the rule verification activity to check each data item according to preset quality rules and verify that the output marking results contain all violation records; if the output quality does not meet the requirements, the rework process is automatically triggered, and the activity is re-executed up to 3 times; if the requirements still cannot be met after repeated processing, the manual intervention mechanism is triggered, and an alarm notification is sent to the data governance manager requesting manual intervention.

[0053] S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit.

[0054] S3 specifically refers to: S3.1. Summarize all output results of each atomic activity, match the template and quality standards in the deliverable storage unit, generate the final deliverable and return it to the user. Summarize the raw dataset output from data query, the violation record markers output from rule validation, the anomaly cause classification output from semantic analysis, and the formatted content output from template matching. Generate a complete PDF report according to the data quality inspection report template and return it to the user. The report template includes a report title, detection time range, detection rule description, violation record details, cause analysis, and improvement suggestion sections.

[0055] S3.2 After the task round is completed and closed, the execution log, deliverables, and user feedback data of this task are synchronously entered into the execution experience memory module's run record storage unit and user feedback storage unit. All execution logs are stored in the relational database of the run record storage unit, the generated PDF report file is stored in object storage with the file path bound to the execution round ID, and user evaluations of the report are stored in the user feedback storage unit.

[0056] S4. The execution log is stored in the running record storage unit, and the user feedback storage unit collects opinions. The task optimization analysis unit periodically collects statistical indicators and triggers the cognitive model memory module (CM) to update.

[0057] S4 specifically refers to: S4.1 The task optimization and analysis unit statistically analyzes operational indicators according to a preset cycle, identifying high-frequency failure nodes, redundant activities, and unreasonable dependencies. It statistically analyzes all task execution records from the past 7 days weekly, calculating the hit rate, delivery accuracy, and average execution time for each task. It identifies rule validation activities with a failure rate as high as 15% when the data volume exceeds 1 million records, thus identifying them as high-frequency failure nodes. It also identifies template matching activities with an average execution time of only 0.1 seconds, accounting for less than 1% of the total execution time, thus identifying them as redundant activities.

[0058] S4.2. Memory update is triggered when any of the following conditions are met: the core metric is below the convergence benchmark for 3 consecutive days, i.e., the delivery accuracy is below 95% for 3 consecutive days; the deliverable fails acceptance or the atomic activity times out, i.e., the report is rejected due to incomplete data or the rule validation activity execution time exceeds the preset 30-second timeout threshold; or the user explicitly proposes a correction, i.e., the user feedback suggests "adding validation rules for null fields".

[0059] S4.3. Generate a targeted optimization proposal, which, after manual review, is used for canary testing with 10% of the traffic. Addressing the high failure rate of rule validation activities under high data volumes, an optimization proposal is generated: "Split the rule validation activity into a batch processing mode, processing 100,000 data entries per batch." After manual review and approval by the data governance manager, the new rules are deployed to 10% of the task traffic for canary testing. After 3 days of canary testing, monitoring shows that the failure rate of the rule validation activity has decreased from 15% to 2%, and the overall task hit rate and delivery accuracy have not decreased. The update is then officially released to all traffic. Simultaneously, the content of the cognitive model memory module before the update is packaged and archived for rollback support. If an anomaly occurs after the update, the system can be restored to the previous version.

[0060] The system and method provided in this embodiment have been practically applied to the data governance scenario of business-finance integration in construction enterprises, providing long-term memory support as a data governance intelligent agent. The cognitive model memory module has five built-in task units: periodic data quality inspection, data quality report generation, problem data query, problem root cause analysis, and governance operation Q&A. Corresponding deliverables include problem data tags, data quality inspection reports, problem data lists, problem analysis reports, and governance operation instructions. Each deliverable corresponds to a set of atomic activities, including: data query to read financial vouchers and contract data from the business-finance database; rule verification to check each of the 88 preset data quality rules; semantic analysis to perform NLP classification and attribution of abnormal data to format errors, logical contradictions, or missing data; template matching to organize and output according to the enterprise management standard format; and report generation to produce an inspection report that conforms to the business-finance integration standard. In actual operation, the main collaborative intelligent agent triggers the periodic data quality inspection task daily at a set time. The task unit storage unit matches the task, retrieves all corresponding atomic activities from the atomic activity storage unit, executes them sequentially, and after verification by the deliverable storage unit, generates a problem data tag deliverable and stores it in object storage. The operation record storage unit retains the full-link execution logs for each inspection, the user feedback storage unit collects the data governance manager's review opinions on the marking results, and the task optimization and analysis unit compiles operational metrics weekly. Operational verification shows that after one month of operation, the task hit rate stabilized at 99%, the delivery accuracy stabilized at 98%, the average delivery time was controlled within 10 seconds, and the manual intervention rate did not exceed 3%.

[0061] The system and method provided in this embodiment can also be applied to other team task scenarios based on software intelligent agents, including general enterprise office team collaboration scenarios. In the document drafting scenario, the task units include document template matching, document content generation, document format verification, and document review. The deliverables include the initial draft of the document, review comments, and the final draft document. The atomic activities include semantic analysis, template filling, rule verification, and format conversion.

Claims

1. A long-term memory system based on a memory-enhanced multi-agent collaborative mental model, characterized in that, The system adopts a binary hierarchical hybrid storage architecture. The system includes a cognitive model memory module (CM), which is connected to an execution experience memory module (EE). The cognitive model memory module (CM) stores the structured cognitive information of team tasks in a three-layer structure of task units, deliverables, and atomic activities. The execution experience memory module (EE) stores the entire process operation data in a hierarchical manner and reverse-optimizes the storage content of the cognitive model memory module (CM).

2. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 1, characterized in that, The cognitive model memory module (CM) includes a task unit storage unit, which stores sub-task information of the team's work after decomposition, the mapping relationship between task unit identifiers and intelligent agents; the sub-task information includes task identifiers, descriptions, status identifiers, task dependencies, and task prompts.

3. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 2, characterized in that, The cognitive model memory module (CM) also includes a deliverable storage unit, which is used to store the standard deliverable information corresponding to each task unit, the mapping relationship between the deliverable identifier and the corresponding task unit identifier; The standard deliverable information includes deliverable identifiers, templates, quality standards, delivery and acceptance entities, and non-conforming iteration rules.

4. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 3, characterized in that, The cognitive model memory module (CM) also includes an atomic activity storage unit, which is used to store the indivisible execution action information corresponding to each deliverable, the mapping relationship between the activity identifier and the corresponding deliverable identifier; The indivisible execution action information includes activity identifier, input / output set, execution rules, execution subject, and agent operation parameters.

5. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 4, characterized in that, The cognitive model memory module (CM) also includes a team member storage unit, which stores information about people and agents participating in the collaboration. The information about people and agents participating in the collaboration includes member identifiers, types, roles, ability descriptions, access permissions, and decision preferences.

6. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 5, characterized in that, The execution experience memory module EE includes a running record storage unit, which is used to retain the full-link execution log of each task and bind it to the corresponding cognitive model entity through an identifier; the full-link execution log includes session information, hit tasks, activity call sequence, time consumption and failure reasons.

7. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 6, characterized in that, The execution experience memory module EE also includes a user feedback storage unit, which is used to collect and associate user evaluations, problem types and specific opinions on each task execution, and automatically bind the corresponding running instance, hit task and deliverable. The Execution Experience Memory Module (EE) also includes a task optimization analysis unit, which is used to periodically calculate task processing volume, hit rate, delivery accuracy, and average time consumption, identify high-frequency failed tasks, redundant activities, and unreasonable processes, and generate optimization proposals.

8. The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 7, characterized in that, The cognitive model memory module (CM) adopts a hybrid storage architecture combining relational databases and vector databases; structured metadata is stored in the relational database; semantic embedding vectors of task descriptions, deliverable names, and activity rules are stored in the vector database; the relational database and the vector database are linked one-to-one by a globally unique identifier. The Execution Experience Memory (EE) module adopts a relational database combined with an object storage architecture; structured execution logs and deliverables are stored in the relational database, while unstructured data of file-type deliverables is stored in object storage; unstructured data is bound to the execution log through task round identifiers.

9. A method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model, characterized in that: The long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 8 comprises: S1. Match the instruction intent and semantics through the task unit storage unit, and combine the decision preferences in the team member storage unit to weighted sort the task to hit; S2. Retrieve all activities through the atomic activity storage unit, arrange the sequence according to dependencies, execute and verify the quality; S3. Generate deliverables and return them to the user by summarizing the activity outputs of the template and quality standards in the deliverable storage unit; S4. The execution log and user feedback storage unit collects opinions, and the task optimization and analysis unit compiles statistics and triggers the cognitive model memory module (CM) to update.

10. The method for operating a long-term memory system based on a memory-enhanced multi-agent collaborative mental model according to claim 9, characterized in that, Specifically, S1 is: S1.1 The main collaborative agent receives natural language instructions sent by the user, completes the task intent and core keyword extraction operation, and filters out noise information. S1.

2. Call the vector database to perform semantic similarity retrieval, recall the Top 10 candidate task units, and perform weighted scoring and sorting according to semantic similarity, recent call frequency, critical path priority and member decision preferences in the team member storage unit; S1.3 When the confidence level of a candidate task is not lower than 0.85, the corresponding task is triggered directly; when the confidence level of a candidate task is lower than 0.6, a request for instruction clarification is sent to the user; when the confidence level is between 0.6 and 0.85, all candidate tasks are displayed for the user to select and confirm. Specifically, S2 is: S2.

1. Locate the unique task unit and its corresponding deliverable, create a task execution round, and record the data of the entire execution process; S2.2 Retrieve all atomic activities corresponding to the deliverable from the atomic activity storage unit, generate an executable activity sequence based on predefined dependencies, and make adaptive fine-tuning of the activity order based on real-time resource usage without violating dependency requirements; S2.

3. Execute each atomic activity sequentially according to the activity sequence, and verify the output quality of each step; if the output quality does not meet the requirements, the rework process is automatically triggered; if the requirements still cannot be met after repeated processing, the manual intervention mechanism is triggered. Specifically, S3 is: S3.1 Summarize all output results of each atomic activity, match the template and quality standards in the deliverable storage unit, generate the final deliverable and return it to the user; S3.2 After the task work round is completed and closed, the execution log, deliverable files and user feedback data of this task are synchronously entered into the running record storage unit and user feedback storage unit in the execution experience memory module. Specifically, S4 is: S4.1 The task optimization analysis unit statistically analyzes operational indicators according to a preset cycle, and identifies high-frequency failure nodes, redundant activities, and unreasonable dependencies. S4.

2. Memory update is triggered when any of the following conditions are met: core metrics are below the convergence benchmark for three consecutive days; deliverables fail acceptance or atomic activities time out; or users explicitly request corrections. S4.3 Generate targeted optimization proposals, which are then manually reviewed and tested with 10% traffic. If no performance degradation is detected, the update is officially released, while retaining historical versions and supporting rollback.

Images