LLM agents get 7 times dumber if they remember their mistakes

When an AI agent tries to fix a mistake, its memory of the failed attempt makes it much worse at the next try. Researchers found that clearing the AI's memory before it retries a task makes it seven times more effective.

People building AI agents often let them retry tasks when they fail. This paper shows that this common practice actually makes the agents perform much worse. Developers now know they must clear the AI's 'mind' after a failure, which means agents can be built to be much more reliable and efficient for complex tasks.