Microsoft Research has found that clearing chat history helps AI agents find better solutions by preventing them from getting stuck in repetitive search patterns. In a new study, researchers examined long-running AI research agents that repeatedly propose, evaluate, and refine solutions based on feedback. The team discovered that these agents can plateau despite receiving additional compute, as accumulated conversation context constrains their search and traps them in specific regions of possible solutions known as "idea basins."
To overcome this limitation, the researchers developed a method called "Fork-and-Flush." This approach splits the agent into parallel branches while wiping their chat contexts clean, though the branches retain their generated files and workspace results. By removing the historical conversation context that steers reasoning back into familiar paths, the agents are able to restore search diversity and explore novel approaches without losing their previous work.
Tested across 13 long-running tasks including cryptanalysis and optimization problems using models via the GitHub Copilot CLI, the Fork-and-Flush method achieved a 66 percent higher normalized average score on the same compute budget compared to single runs. The technique also produced more consistent results, demonstrating that regular context resets are crucial for improving the performance and exploration capabilities of autonomous research agents.
Why it matters: This research demonstrates that managing conversation context through strategic resets can significantly boost AI agent performance on complex tasks without requiring additional compute resources.