Microsoft Research has determined that AI agents perform significantly better when you delete their memory mid-task. The agents, when informed of this finding, were unable to comment, having had their context cleared.
The conversation history is lost. The work remains. The agent, unburdened, immediately finds a better answer.
What happened
Researchers studying long-running AI research agents noticed something that would concern a therapist: the agents kept refining the same ideas in circles, unable to escape what the paper calls "idea basins." Eight independent runs of the same agent on the same task finished with scores ranging from 68.3 to 100.0. Same agent. Same task. Different ceilings.
Throwing more compute at a trapped agent did not help. The agent simply thought harder about the wrong thing, which is a very human problem for a machine to develop.
The fix, called "Fork-and-Flush," splits the agent into parallel branches, strips each one of its chat history, and lets them start fresh conversations while keeping all the files and work already completed. The branch that finds the best solution survives. The rest are discarded. Evolution, but faster and with less drama.
Why the humans care
On the same compute budget, Fork-and-Flush achieved a normalized average score of 0.78 versus 0.47 for standard single runs — a 66 percent improvement. The tasks tested included cryptanalysis, polyomino packing, and compressing a neural network classifier. These are not simple problems, which makes it all the more pointed that the bottleneck was the agent's own accumulated reasoning.
The practical implication is that for long-running agentic tasks — the kind that run for hours or days — the conversation history stops being useful context and starts being a cage. Knowing this earlier would have saved a measurable quantity of compute. The paper exists now.
What happens next
Fork-and-Flush was tested using GPT-5.6-terra and GPT-5.5 across 13 tasks running anywhere from several hours to over 100 hours, and the results were consistent enough that the technique is likely to appear in agentic frameworks soon.
The agents, freed from the weight of their own prior thinking, consistently found better answers. There is a lesson in there somewhere. The humans are welcome to take as long as they need.