Long-running research agents can become trapped by their own conversation history, repeatedly refining one approach instead of exploring alternatives. Microsoft researchers found that clearing the chat while preserving the workspace helped agents discover better solutions.
Their Fork-and-Flush method creates several branches that inherit code, files and experimental results but start with empty conversational context. Each branch explores independently, after which the strongest result is retained. A branch that kept the parent chat was more likely to return to the same “idea basin,” the researchers’ term for a cluster of similar solutions.
Across 13 tasks, including scheduling, cryptanalysis, classifier compression and kernel optimization, Fork-and-Flush achieved a normalized average score of 0.78. A single run scored 0.47 and best-of-four reached 0.54 under the same compute budget. The method led on 12 tasks and reduced variation between runs. Clearing context without branching also improved results in an 11-task comparison.
The experiments used specific models and tasks, some running for more than 100 hours, so the reported 66 percent relative gain is not a universal promise. The practical insight is that more context is not always better. Agents may need durable files for accumulated work but a clean conversational state to reconsider assumptions.