Clearing chat history helps AI agents find better solutions
decoder ·
A Microsoft Research study reveals that long-running AI agents often plateau because accumulated conversation history constrains their search within narrow 'idea basins.' To address this, researchers developed a method called 'Fork-and-Flush,' which splits the agent into parallel branches while clearing the chat context to encourage exploration of new solutions. This approach allows agents to discard previous reasoning biases while retaining generated files and code in the workspace. Testing on 13 complex tasks showed that Fork-and-Flush achieved a 66 percent higher normalized average score compared to single runs with the same compute budget. The method also demonstrated greater consistency, reducing the standard deviation of results significantly. Even simple context resets without branching improved performance by 22 percent. Researchers note that as models become more capable, the variation caused by these resets will likely persist at the frontier of problem-solving difficulty. Consequently, regular context resets are expected to remain a valuable strategy for future autonomous research agents.
decoder