Multi-Agent Transactive Memory: shared agent trajectories

Multi-Agent Transactive Memory (MATM), submitted to arXiv on 18 Jun 2026, is a framework for population-level storage and retrieval of agent-generated trajectories that lets agents share reusable procedural knowledge. The paper argues producer agents contribute trajectories to a shared repository and consumer agents retrieve them to improve task execution.

What is Multi-Agent Transactive Memory?

Multi-Agent Transactive Memory is a retrieval-based system for agent populations: producer agents contribute trajectories to a shared repository and consumer agents retrieve those trajectories to support task execution. The authors frame MATM as an extension of retrieval-augmented generation to agent-generated artifacts, aiming to prevent repeated rediscovery of solutions by newly instantiated agents.

The paper positions trajectories as reusable procedural knowledge, and treats retrieval systems as the analogue to search engines that index human artifacts for human problem solving. Authors include To Eun Kim, Xuhong He, Dishank Jain, Ambuj Agrawal, Negar Arabzadeh, and Fernando Diaz.

How did the authors evaluate MATM?

The authors evaluated MATM in interactive environments where trajectories are long and encode rich procedural structure, specifically ALFWorld and WebArena. They run experiments that let producer agents add trajectories to the repository and let consumer agents retrieve them during task execution.

The paper reports that "retrieving trajectories from MATM improves downstream task performance and reduces interaction steps without coordination or joint training." The experiments therefore test whether population-level retrieval of agent-generated artifacts helps agents complete tasks more effectively and with fewer interactions. The environments named in the study are ALFWorld and WebArena, which the authors cite as contexts with especially rich procedural trajectories.

Why it matters

MATM tackles a practical inefficiency: agent-generated procedural traces are often discarded or siloed inside the producing agent, forcing others to rediscover solutions. By centralizing and indexing trajectories, MATM enables reuse across heterogeneous agents and tasks, reducing redundant search and interaction. The approach requires no coordination or joint training between agents, which lowers the integration barrier for open, decentralized agent ecosystems.

This matters for systems that deploy many specialized agents across diverse tasks, because it offers a concrete design pattern for experience sharing: persistent, searchable trajectories that other agents can retrieve and apply.

What to watch

Look for follow-up work that provides quantitative metrics on how MATM scales with larger repositories and more diverse agent populations, and for code or dataset releases tied to the ALFWorld and WebArena experiments. The paper itself frames MATM as a design pattern for population-level experience sharing in open agent ecosystems, so adoption or replication by other teams will be the clearest signal of broader impact.

References and key facts from the paper: submission date 18 Jun 2026; environments used: ALFWorld and WebArena; authors: To Eun Kim, Xuhong He, Dishank Jain, Ambuj Agrawal, Negar Arabzadeh, Fernando Diaz; core claim: retrieving trajectories from MATM improves downstream task performance and reduces interaction steps without coordination or joint training.