For open-ended discovery tasks, strategically forgetting accumulated history is as important as raw compute—agents that periodically reset their context but retain best solutions explore more effectively than those that remember everything.
This paper shows that LLM agents searching for solutions hit a plateau when they accumulate too much history, becoming anchored to past attempts. The key insight is that periodically forgetting the search history and restarting from the best solution found so far allows agents to explore more effectively.