Berkeley/Stanford paper: agents gain up to 39% by wiping memory and restarting from best result
alexcovo_eth · x · 2026-11-01
A paper from UC Berkeley, Stanford and Bespoke Labs ("Scaling Discovery Requires Forgetting", arXiv 2610.20563) shows agents in long research/optimization loops plateau because accumulated history makes them anchor on old attempts and shrink their edits.
The fix is simple:
- wipe the session and candidate history each round
- keep only the best result so far
- restart fresh from it, and repeat
Same model, same token budget, across 4 discovery tasks: up to 39% better than a single agent and up to 15% better than parallel agents. The lesson: keep the artifact, drop the transcript.
More from coding & agent
- Hands-on tutorial trains a decision model on Qwen3-0.6B from scratch — dl_weekly · 2026-11-01
- ui-ux-pro-max and 6 more repos giving AI coding tools real design taste — GCWebDesigner · 2026-11-01
- Hex ships stable release: local voice dictation for talking to AI agents on Mac, Windows, Linux — madeofAjala · 2026-11-01
- Personal AI agent race heats up — this prompt migrates your entire setup between agents — EXM7777 · 2026-11-01
- Nous Research's Hermes Agent launches mobile beta on iOS and Android via TestFlight — testingcatalog · 2026-11-01
- How to Build a Production AI Agent: The 7 Layers Between the Model and the Result — blaizedsouza · 2026-11-01