An agent can retain every interaction and still repeat its mistakes. The useful unit of memory here is an actionable lesson, including what went wrong, rather than an entire transcript. That makes the paper relevant to systems that repeatedly encounter related tasks.
Key Findings
- On WebArena with Gemini 2.5 Flash, success rises from 40.5% without memory to 48.8%, while mean interaction steps fall from 9.7 to 8.3. This measures reuse across a task stream, rather than model weight updates.
- Lessons come from both successes and failures. Self-judged outcomes supply strategies and warnings without requiring ground-truth feedback during learning.
- MaTTS compares multiple attempts or refines one attempt before extracting memories. Extra computation can therefore improve later tasks as well as the current answer. The authors release implementations for web browsing and software engineering.
Who Should Read It
Great fit if you are designing agents that should improve across recurring tasks, especially browser or coding workflows. Look elsewhere if you need a production-ready memory service: the released code is a research demonstration, and judge errors, retrieval quality, and added inference work still need evaluation in your environment.