| 指標 | 數值 | 說明 |
|---|---|---|
| Test environments | ALFWorld + WebArena | Two standard interactive agent benchmarks with long, procedurally-rich trajectories |
| Performance gain | Significant improvement | Consumer agents retrieving from MATM outperform no-memory baselines on downstream tasks |
| Interaction steps | Reduced | Agents with trajectory retrieval need fewer steps to complete tasks |
| Coordination needed | Zero | No inter-agent communication or joint training required; pure passive retrieval |
| Author affiliations | UC Berkeley + CMU + Waterloo | Fernando Diaz (Google DeepMind) served as advisor |
After completing a task, stores the full (state, action, observation, reward) trajectory into the shared MATM repository
When facing a new task, retrieves similar historical trajectories from MATM and injects them as few-shot examples into its prompt
Converts trajectories into dense vectors for semantic similarity search - like RAG's retriever but operating on agent trajectories instead of human documents
Population-level persistent storage; all agents read and write. Design pattern analogous to a human organization's knowledge base
If each Hermes profile's skill execution trajectories were stored in a shared repository, cross-profile agents could learn from each other's experiences. For example, the tools profile's deployment tricks could be retrieved by the learn profile without re-exploration.
Complex case trajectories solved by each agent go into MATM; newly deployed agents facing similar issues immediately have reference examples, drastically reducing onboarding cost.
If MATM becomes an open standard, developers could share agent trajectory datasets, similar to HuggingFace for model weights. This shifts the agent learning curve from 'every agent starts from zero' to 'standing on the shoulders of the entire community.'
MATM's core insight: agent trajectories are not disposable waste but reusable procedural knowledge assets. This has direct implications for Hermes architecture: our current memory module only stores 'results' (summaries), but the most valuable part is the 'path' (what I did, what I tried, where I failed). If we could log complete tool-call trajectories (not just final outputs), cross-session learning efficiency would dramatically improve. This paper gives a clear engineering direction: add trajectory-level logging to Hermes memory with cross-profile retrieval support.
📎 論文連結：arxiv.org/abs/2606.19911