Microsoft researchers find clearing chat history helps AI agents find better solutions

Microsoft Research reports that long-running AI research agents can plateau in a narrow "idea basin" despite more compute.

The proposed Fork-and-Flush method runs parallel branches with fresh chat contexts while keeping the workspace, reaching a 0.78 normalized average score versus 0.47 for single runs on 13 tasks.

The authors say regular context resets could remain useful even for frontier research agents.