A research framework proposes letting AI agents pool their experience so newly spun-up agents stop re-solving problems that other agents already solved.
The paper introduces Multi-Agent Transactive Memory (MATM), a design pattern in which agents contribute their completed task trajectories - the step-by-step records of how they navigated a problem - to a shared repository. Other agents can then retrieve those trajectories and use them as procedural guides. The researchers tested the approach in two interactive benchmarks, ALFWorld and WebArena, where tasks require long action sequences. Agents using MATM completed tasks in fewer steps and with higher success rates than agents working from scratch, with no joint training or real-time coordination required.
The underlying observation is mundane but the implication isn't: every time an AI agent figures out how to do something, that knowledge currently dies with the session. MATM treats agent-generated workflows the way search engines treat web pages - as indexable, retrievable artifacts with ongoing utility. For teams running large fleets of task-specific agents, that shift from ephemeral to persistent procedural knowledge could meaningfully reduce compute overhead.
The comparison the paper draws to search engines is apt, but it also flags the obvious next problem: garbage-in, garbage-out. If agents contribute bad trajectories to the shared pool, retrieval makes things worse, not better - a quality-control challenge the research does not yet fully address.