A provenance-aware selective memory sharing method for multiagent decision-making

Abstract

As multi-agent systems are increasingly used to handle complex tasks, procedures such as retrieval, filtering, reasoning, and explanation are often assigned to different agents. External memory, shared memory, and communication mechanisms have therefore become key factors shaping collaborative efficiency. In many complex decision-making tasks, each agent first analyzes the local memories available to it, and the agents then jointly form the final judgment. Studies such as SimpleMem and AgeMem mainly focus on long-term memory compression and multi-granularity memory retrieval. In multi-agent systems, however, memory storage is only one side of the problem; how memory should be shared is equally important. Broadcasting all candidate memories among agents can reduce the risk of missing critical evidence, but it also increases communication cost, contextual redundancy, and noise. By contrast, sharing no memory, or sharing only according to a fixed similarity threshold, may miss key clues required for cross-source reasoning. To address these issues, this paper proposes a provenance-aware selective memory sharing method, namely Provenance-Aware Memory Selective Sharing (PA-MemShare). The method formulates candidate memory sharing as a lightweight policy learning problem and introduces a Provenance-Aware Memory Credit Assignment mechanism (PMCA), which uses source documents and sentence indices for weakly supervised attribution. Experiments use the HotpotQA multi-hop question answering dataset as a proxy benchmark for knowledge-intensive multi-agent decision-making tasks. The results show that PA-MemShare achieves 66.48% F1 and 47.72% SF F1 on the test set, with an average Token Cost of 80.6. Further comparison experiments indicate that the proposed method can improve shared evidence quality while keeping communication cost under control.