Multi-agent cooperation method and device, electronic equipment and storage medium

By introducing process memory and summary memory into a multi-agent system, the problems of narrow and redundant memory strategies in multi-agent collaboration are solved, enabling precise adaptation and efficient invocation of agents, and improving the system's synergy and user problem understanding capabilities.

CN122309191APending Publication Date: 2026-06-30中国邮政储蓄银行股份有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
中国邮政储蓄银行股份有限公司
Filing Date
2026-03-05
Publication Date
2026-06-30

AI Technical Summary

Technical Problem

In existing multi-agent systems, the agent memory strategy has become a key constraint during collaboration. The multi-agent memory components are flat and have a narrow scope, making it difficult to effectively handle complex information such as sequential events and emotional states. This leads to poor performance in long-term tasks and in handling user-specific information. Furthermore, multiple rounds of interaction may result in an excessive number of input tokens for the request model.

Method used

A dual-memory system is adopted, including process memory and summary memory. Process memory records the timing trajectory and triggering conditions of agent calls, while summary memory is used to simplify the information of each round of interaction. The range of agents to be called is determined through the model decision-making stage, and the generated content is decomposed in the generation stage. The dual-memory system is combined to achieve accurate adaptation and efficient calling of agents.

Benefits of technology

Through the collaborative mechanism of the dual-memory system, automated and rational scheduling in the multi-agent collaboration process is achieved, enhancing the ability to understand and decompose user fuzzy problems, reducing memory redundancy, and improving the collaboration and efficiency of the multi-agent system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122309191A_ABST
    Figure CN122309191A_ABST
Patent Text Reader

Abstract

This application discloses a multi-agent collaboration method, apparatus, electronic device, and storage medium. The method includes, in response to a user-input query, determining the scope of agents to be invoked in a model decision-making stage. This model decision-making stage invokes at least process memory and summary memory. The process memory records the temporal trajectory, type, and triggering conditions of each agent's invocation, while the summary memory simplifies the memory information for each round of interaction. Based on the scope of agents to be invoked, the generated content of each agent is decomposed in a model generation stage, which invokes at least the summary memory. Finally, based on the generated content of each agent, a final output result is obtained. This application achieves a stronger understanding and decomposition capability for user-fuzzy questions, understanding user questions at the memory level, thereby invoking agents not mentioned in the user question and completing multi-agent collaboration.
Need to check novelty before this filing date? Find Prior Art