Multi-Agent Transactive Memory

The article introduces Multi-Agent Transactive Memory (MATM), a framework designed for the population-level storage and retrieval of agent-generated trajectories to enhance knowledge sharing among diverse LLM agents. By allowing producer agents to contribute their trajectories to a shared repository, MATM enables consumer agents to retrieve these artifacts, improving task execution in interactive environments like ALFWorld and WebArena. The experimental results indicate that using MATM significantly enhances downstream task performance and reduces interaction steps, highlighting its potential as a design pattern for experience sharing in decentralized agent ecosystems.