Microsoft study: clearing chat history helps AI research agents find better solutions
A Microsoft Research study finds that long-running AI research agents plateau because accumulated conversation context traps them in narrow "idea basins" within the solution space. In eight independent runs of the same agent on the same task, final scores ranged from 68.3 to 100.0, and the gaps persisted even with substantially more compute. The proposed Fork-and-Flush method splits the agent into parallel branches that inherit the workspace and files but lose all chat context; across 13 long-running tasks (including polyomino packing, job-shop scheduling and cryptanalysis) it reached a min-max-normalized average score of 0.78 versus 0.47 for single runs and 0.54 for best-of-4 — relative gains of 66.0% and 44.4% — with normalized standard deviation falling from 0.26 to 0.16. Flushing context without branching alone still improved results by about 22%; the agents used GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI, with individual runs lasting hours to over 100 hours.
If the finding holds, frontier research agents may still need periodic context resets rather than relying on more compute or stronger models to escape stalled searches.