MATM Framework: Multi-Agent Transactive Memory Enables Population-Level Trajectory Sharing

841io · x · 2026-08-28

Introduces Multi-Agent Transactive Memory (MATM), a framework for sharing knowledge across heterogeneous LLM agent populations. It extends RAG from human-authored artifacts to agent-generated artifacts, specifically task trajectories. Producers contribute trajectories to a shared repository, which consumers retrieve to improve task execution. Experiments on ALFWorld and WebArena show that retrieving trajectories from MATM improves downstream performance and reduces interaction steps without requiring coordination or joint training, establishing it as a design pattern for experience sharing in open agent ecosystems.

Original post →

More from coding & agent

coding & agent channel →