Multi-Agent Transactive Memory Enhances LLM Agent Collaboration

The increasing deployment of Large Language Model (LLM) agents with diverse capabilities across various tasks highlights the need for effective knowledge sharing among these heterogeneous populations. Similar to how search engines organize human-generated content, retrieval systems can now organize agent-generated artifacts for reuse. Specifically, agent trajectories, which encode valuable procedural knowledge, are often discarded after a single use or kept only by the originating agent. This forces new agents to repeatedly rediscover solutions that already exist. To address this inefficiency, the authors propose Multi-Agent Transactive Memory (MATM). MATM is a framework designed for population-level storage and retrieval of these agent-generated trajectories. Producer agents contribute their task execution paths to a shared repository, while consumer agents can retrieve these trajectories to enhance their own task performance. Focusing on interactive environments like ALFWorld and WebArena, where trajectories are long and rich in procedural structure, experiments demonstrate that retrieving from MATM significantly improves downstream task performance and reduces the number of interaction steps. This is achieved without requiring explicit coordination or joint training among agents, positioning MATM as a valuable design pattern for experience sharing in open agent ecosystems.