MATM framework lets AI agents share procedural memory across populations
Researchers propose Multi-Agent Transactive Memory (MATM), a framework that stores and retrieves agent-generated trajectories in a shared repository so that newly instantiated agents can reuse procedural knowledge instead of rediscovering solutions from scratch.
Score breakdown
MATM removes the repeated rediscovery cost baked into stateless agent deployments by giving heterogeneous agent populations a shared, retrievable store of procedural experience — without requiring joint training or inter-agent coordination.
- 01Proposed by To Eun Kim, Xuhong He, and Dishank Jain (arXiv, 2026-06-18)
- 02MATM stands for Multi-Agent Transactive Memory, a framework for population-level storage and retrieval of agent-generated trajectories
- 03Agent trajectories are typically discarded after a single use or retained only by the producing agent, motivating the shared repository design
To Eun Kim, Xuhong He, and Dishank Jain introduce Multi-Agent Transactive Memory (MATM), a framework designed to address a fundamental inefficiency in decentralized LLM agent deployments: agent-generated trajectories — which encode rich procedural knowledge — are typically discarded after a single use or kept only by the producing agent, forcing every newly instantiated agent to rediscover existing solutions independently. MATM draws an analogy to search engines indexing human-generated artifacts, extending retrieval-augmented generation (RAG) from individual agents consuming human-authored content to entire agent populations consuming agent-authored trajectories. Producer agents contribute their task trajectories to a shared repository, and consumer agents retrieve relevant trajectories to guide their own task execution.
The authors evaluate MATM in interactive environments — ALFWorld and WebArena — chosen because their trajectories are long and encode especially rich procedural structure.
The authors evaluate MATM in interactive environments — ALFWorld and WebArena — chosen because their trajectories are long and encode especially rich procedural structure. Results show that retrieving trajectories from MATM improves downstream task performance and reduces the number of interaction steps required, without any coordination mechanism or joint training across agents. The paper frames MATM as a design pattern for population-level experience sharing in open agent ecosystems.
Key facts
- 01Proposed by To Eun Kim, Xuhong He, and Dishank Jain (arXiv, 2026-06-18)
- 02MATM stands for Multi-Agent Transactive Memory, a framework for population-level storage and retrieval of agent-generated trajectories
- 03Agent trajectories are typically discarded after a single use or retained only by the producing agent, motivating the shared repository design
- 04Producer agents contribute trajectories to a shared repository; consumer agents retrieve them to improve task execution
- 05Evaluated on ALFWorld and WebArena interactive environments, where trajectories are long and encode rich procedural structure
- 06MATM improves downstream task performance and reduces interaction steps without coordination or joint training
- 07The authors frame MATM as a design pattern for population-level experience sharing in open agent ecosystems
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →