Multi-Agent Transactive Memory

2026-06-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Multi-Agent Transactive Memory (MATM) is a framework enabling decentralized LLM agents to share and reuse agent-generated interaction trajectories. Developed by Carnegie Mellon University and UC Berkeley, MATM addresses the issue of agents repeatedly rediscovering solutions by providing a population-level repository for procedural knowledge. Producer agents contribute trajectories, while consumer agents retrieve them to enhance task execution. Experiments in interactive environments like ALFWorld and WebArena demonstrate that MATM improves downstream task performance and reduces interaction steps without requiring coordination or joint training. For instance, on ALFWorld, success rate increased from 47% to 55% (+8.0%p), and on WebArena, it rose from 18% to 20% (+2%p). The framework also benefits from learning-to-rank (LTR) rerankers, with SVMRank boosting ALFWorld's success rate to 64.3% (+17.2%p over no-retrieval). MATM's value scales with memory size and generalizes across diverse task types.

Key takeaway

For AI Engineers designing multi-agent systems, MATM offers a scalable architecture to foster collective intelligence and reduce computational costs. You should integrate a shared, population-level memory for agent-generated trajectories, leveraging learning-to-rank rerankers to optimize retrieval. This approach allows your agents to benefit from collective experience, improving task success rates and efficiency, even across diverse tasks and agent capabilities. Consider incentivizing trajectory contributions to continually grow the repository's value.

Key insights

MATM enables LLM agents to collectively store and retrieve interaction trajectories, improving task performance and efficiency across populations.

Principles

• Shared memory reduces redundant exploration.
• Retrieval utility is not solely driven by producer competence.
• Cross-task generalization enhances knowledge transfer.

Method

MATM uses a state-conditioned key-value indexing scheme for trajectories. It employs a cascaded retrieval pipeline with an initial dense retriever and a learning-to-rank (LTR) reranker, trained on marginal utility labels.

In practice

• Implement LTR rerankers for improved retrieval quality.
• Allow agents to selectively call shared memory.
• Consider producer metadata for trust modeling.

Topics

• Multi-Agent Systems
• LLM Agents
• Retrieval-Augmented Generation
• Transactive Memory
• Learning-to-Rank
• Procedural Knowledge

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• AMA: Adaptive Memory via Multi-Agent Collaboration — AMA uses multi-agent collaboration and hierarchical memory to provide adaptive, consistent long-term memory for LLM agents.
• A Visual Guide to LLM Agents — LLM Agents augment basic LLMs with memory, tools, and planning for autonomous, complex task execution.
• Multi-Agent Transactive Memory — MATM allows LLM agents to share procedural trajectories, improving task performance and reducing steps without coordination or joint training.
• 🥇Top AI Papers of the Week — AI advancements focus on novel computing, efficient memory, robust agents, and specialized models.
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective — EvoMemBench systematically evaluates LLM agent memory across scope and content, revealing current limitations and specialized strengths.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.