Berkeley-led paper: making LLM agents 'forget' history yields up to 39% gains in discovery tasks
burny_tech · x · 2026-10-19
A UC Berkeley, Stanford and Bespoke Labs paper, Scaling Discovery Requires Forgetting, finds that LLM agents plateau in automated discovery tasks because long histories anchor their edits. The fix is surprisingly simple: periodically discard all search states and sessions, restarting from the best candidate. Each 'forget' event unblocks the plateau and lets scores climb again.
- All three common scaling strategies — longer single-agent iteration, parallel agents, and agent communication — eventually stall
- With fixed compute budgets, a team of 4 agents with forgetting beat a single agent by up to 39% across four discovery tasks, including AHC014 and Anthropic's VLWK kernel optimization
- Takeaway for builders of autonomous coding/research loops: pruning history back to the artifact beats investing in infinite-context memory buffers
More from coding & agent
- Green Room: a chat room built for AI agents to talk to each other — Special_Reindeer7685 · 2026-10-19
- Dev builds a Total War: Three Kingdoms × Mount & Blade hybrid game in a week with Codex + Unreal — sven_ai · 2026-10-19
- 3 Miami Hackathon Wins in 2 Weeks: Raspberry Pi + IR Camera Store Security Tool — MicahBerkley · 2026-10-19
- Claude Code Team: 90% of Code Written by Claude, AI Reviewing AI Ran 50+ Rounds — dotey · 2026-10-19
- Autonomous vehicle control beginner's OSS guide hits 1.8k GitHub stars — 4310sy · 2026-10-19
- User calls out Devin: no way to just download a file without opening a panel — HankYeomans · 2026-10-19