Multi-Agent Transactive Memory: Sharing Agent Trajectories Boosts Task Performance

mrdrozdov · x · 2026-09-22

mrdrozdov highlights two multi-agent papers, chiefly the arXiv paper "Multi-Agent Transactive Memory" (MATM). As heterogeneous LLM agents are deployed across diverse tasks, infrastructure for knowledge sharing across agent populations becomes necessary. MATM extends RAG from retrieving human-authored artifacts for individual agents to retrieving agent-generated artifacts for a population: producer agents contribute trajectories (encoding reusable procedural knowledge) to a shared repository, and consumer agents retrieve them to improve execution, avoiding repeated rediscovery of existing solutions.

Experiments on ALFWorld and WebArena — environments with long, procedurally rich trajectories — show MATM retrieval improves downstream performance and reduces interaction steps without coordination or joint training, positioning it as a design pattern for population-level experience sharing in open agent ecosystems. Also recommended: the 2025 "LLM-Based Multi-Agent Blackboard System for Information Discovery in Data Science."

Original post →

More from coding & agent

coding & agent channel →