Microsoft researchers find clearing chat history improves AI agents

Long-running AI research agents can get stuck refining solutions within limited problem spaces, called "idea basins," even when given more computing power to work with. Microsoft Research found that clearing an agent's conversation history while preserving its saved work allows it to explore different solution approaches and achieve significantly better results. Their method, called "Fork-and-Flush," creates parallel branches of an agent with fresh chat contexts while keeping the underlying workspace, resulting in 66 percent higher average scores on the same computing budget compared to single runs. The researchers suggest that even as AI models become more capable, resetting context may remain useful for tackling open research problems that sit at the limits of what models can do.