Shared agent memories boost performance without coordination

The authors propose Multi-Agent Transactive Memory, a repository for agent-generated trajectories that improves task success and reduces steps in interactive environments.

PaperTop Universitycs.AIarXiv:2606.19911v1 ↗

To Eun Kim · Xuhong He · Dishank Jain · Ambuj Agrawal · Negar Arabzadeh · Fernando Diaz

Carnegie Mellon University · University of California, Berkeley

Research Digest·18 June 2026·2 min read

Read the paper →

Thread:Memory Management for Agents

What they did

The authors developed MATM, where producer agents contribute trajectories—sequences of actions and observations—to a shared vector database. Consumer agents retrieve relevant trajectories using semantic similarity to the current task prompt and execute the most similar one. The framework was evaluated on ALFWorld, a text-based household task simulator, and WebArena, a web-based task environment. Experiments involved three LLM backbones (GPT-4, GPT-4o, and GPT-4o-mini) and used zero-shot baselines, with retrieval from trajectories generated by agents of the same or different models.

Key findings

MATM improved task success rates by 2.5% to 24% over zero-shot baselines across environments and models, with the largest gains in WebArena using GPT-4o-mini (from 16.9% to 41.0%).
Retrieved trajectories reduced average interaction steps by 21–37% for successful tasks in ALFWorld, indicating more efficient execution.
Trajectories from agents with different backbones (e.g., GPT-4 aiding GPT-4o-mini) still improved performance, showing cross-model transfer is viable.
MATM outperformed a naive memory baseline (retrieving random trajectories) and was robust to up to 40% corruption in the repository.

Why it matters

Current LLM agent deployments discard or silo trajectory data, forcing each new agent to rediscover solutions. MATM demonstrates that population-level memory can be achieved without explicit coordination, joint training, or handcrafted pipelines, offering a scalable design pattern for open agent ecosystems where agents are created and managed by different parties.

Caveats

The experiments are limited to two simulated interactive environments (ALFWorld and WebArena) with relatively short trajectories; generalizability to more complex or long-horizon tasks is untested. The retrieved trajectory is used as a direct prompt, which may risk overfitting or propagation of suboptimal strategies if the repository contains low-quality trajectories. The study also assumes a stable population and does not address memory freshness or concept drift over time.