Clearing chat history boosts AI agent solutions

microsoft research found that long-running ai agents can plateau despite additional compute, with accumulated conversation context constraining their search. a study on autoresearch agents showed that eight independent runs on the same task plateaued at different performance levels, ranging from 68.3 to 100.0. the fork-and-flush method, which combines forking and flushing, improved scores by 66% on 13 long-running tasks. this method achieved a min-max-normalized average score of 0.78, compared to 0.47 for single runs.