Microsoft Research finds clearing chat history boosts AI agent performance by 66%
Microsoft Research published a study showing that long-running AI agents often plateau in performance despite more compute. The issue stems from accumulated conversation context, which restricts agents to similar solutions in what researchers call idea basins—regions of the solution space where they repeatedly refine the same approaches without exploring better ones.
Key points
- Microsoft Research shows AI agents plateau due to accumulated chat context, limiting solution diversity
- *Fork-and-Flush* method boosts agent performance by 66% on 13 tasks by clearing chat history while preserving workspace
- Even weaker models (Qwen3.5-4B to 27B) show performance variance on hard problems, hinting at persistent need for context resets
The team tested a method called Fork-and-Flush, which splits an agent into parallel branches and clears chat history while preserving workspace files. Even clearing chat context without branching yielded a 22% gain. The study used GPT-5.6-terra and GPT-5.5 via GitHub Copilot CLI, with individual runs lasting hours to over 100 hours.
Researchers also found that stronger models (Qwen3.5-4B, 9B, 27B) still face variability in performance on hard problems, suggesting context resets may remain useful as models advance.
Model page: GPT-5.6 Terra →
The story so far
2 episodes →- Microsoft Research finds clearing chat history boosts AI agent performance by 66%this story
Clearing chat history helps AI agents find better solutions
The Decoder · 12 October 2026
Loading the full article…
This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic's Claude begins 'dreaming' to reorganize memory in tests · 1 src
- John Langford quits Microsoft, declares AI Frontiers over · 1 src
- Anthropic releases study on Claude usage in scientific research · 3 src
- MIT and BASF release CheMeleon for low-data molecular prediction · 2 src
- Anthropic tests Claude agents in EVE Online trading and sabotage detection · 1 src
Comments
via GitHub Discussions