Clearing chat history helps AI agents find better solutions
Microsoft Research shows long-running AI agents plateau in "idea basins"; a Fork-and-Flush context-reset method boosts scores 66% on equal compute.
Microsoft researchers found that "Autoresearch agents" get trapped in "idea basins," refining similar solutions despite added compute, with identical tasks yielding final scores from 68.3 to 100.0 across runs. Their Fork-and-Flush method spawns K parallel branches that flush chat history but keep the workspace, reaching a normalized average score of 0.78 versus 0.47 for single runs across 13 long-running tasks (a 66% relative gain). Tests used GPT-5.6-terra and GPT-5.5 via GitHub Copilot CLI, and variance across Qwen3.5-4B/9B/27B sizes peaked on tasks at the edge of model capability.
48