Microsoft Research: Clearing Chat History Helps AI Agents Escape Stuck Solution Paths

A new Microsoft Research study shows that long-running AI research agents can plateau at strikingly different performance levels on identical tasks, and that their own accumulated conversation history may be the cause. The researchers found that context buildup pushes agents into what they call "idea basins," narrow regions of similar solutions, and that resetting the chat history can unlock substantially better results.

Agents Get Stuck in Idea Basins

The researchers studied "Autoresearch agents," AI systems that repeatedly propose solutions, evaluate them, and refine them based on feedback. Eight independent runs of the same agent on the same task plateaued at significantly different levels, with final scores ranging from 68.3 to 100.0, and the gaps persisted even with substantially more compute.

By plotting solutions from all runs on a two-dimensional map where similar solutions sit close together, the team found that solutions from a single run almost always stayed within a small region. The agents then spend a large share of their remaining compute budget refining solutions inside a less productive region instead of finding the better approaches discovered by other runs. This pattern persists even when the agents periodically compress their conversation histories.

Fork-and-Flush Clears the Context

The proposed method combines two mechanisms. In "forking," the agent splits into K parallel branches, each inheriting the previous workspace and all generated files. In "flushing," each branch loses its entire chat context but keeps the workspace, allowing it to start a fresh conversation and build on prior work.

A branch that keeps the parent process's chat context tends to return to the same idea basin, because the previous reasoning steers its next steps. Fresh context restores diversity in the search. The conversation history is lost, but the code, files, and saved results remain in the workspace.

Results on 13 Long-Running Tasks

Fork-and-Flush was tested on 13 long-running tasks, including polyomino packing, job-shop scheduling, cryptanalysis, compressing a CIFAR classifier, and optimizing a VLIW kernel. Individual runs lasted from several hours to over 100 hours, and the agents used GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI.

On the same compute budget, Fork-and-Flush achieved a min-max-normalized average score of 0.78, compared with 0.47 for single runs and 0.54 for best-of-4, relative gains of 66.0 and 44.4 percent. The method achieved the highest mean score on 12 of the 13 tasks and produced more consistent results, with the normalized standard deviation falling to 0.16 from 0.26 for single runs. Simply clearing the chat context without branching improved results by about 22 percent on 11 tasks.

Why Context Resets May Stay Useful

The researchers also tested whether more capable models avoid idea basins, using MAX-3-SAT problems across three model sizes, Qwen3.5-4B, 9B, and 27B. Variance between runs followed an inverted U-shaped curve, peaking on tasks at the edge of a model's abilities, roughly halfway between random-guessing performance and the reference solver's performance. Models almost always succeed on very easy tasks and fail on very hard ones.

The authors argue that open research problems sit, by definition, at the limits of models' abilities. They expect the variation Fork-and-Flush exploits to persist as models improve, with the boundary simply shifting toward harder problems, and they conclude that regular context resets could remain useful even for frontier research agents.

Sources: The Decoder