EMSP: enhancing multi-agent service performance based on historical behavior memory

With the increasing application of Large Language Models (LLMs) in service composition, the paradigm of multi-agent collaboration has emerged as a critical approach for tackling complex tasks that require efficient orchestration of various services. Despite its promise, practical implementations often encounter significant challenges in enhancing the efficiency of agent interactions, minimizing operational errors, and optimizing the process of service invocation. In this paper, we propos e a novel historically reinforced Prompt enhancement mechanism that strategically leverages interactions between agents and user feedback to bolster the capabilities of agents with diverse functionalities. This mechanism involves the evaluation of agent performance through tailored reward functions, as well as the incorporation of memor y reinforcement strategies that effectively preserve and optimize beneficial behaviors over time. As a result, we observe a gradual enhancement in the qualit y of agent-generated Prompts. Comprehensive experimental results reveal marked improvements in key performance metrics, including planning accuracy and parameter accuracy, alongside a notable reduction in hallucination rates across various API composition tasks. These findings collectively underscore the efficacy of our proposed approach in optimizing the service invocation process for complex computational challenges, paving the way for more reliable and intelligent service composition frameworks.