Microsoft Wipes AI Agents' Chat History and Gets Up to 66% Better Results on the Same Budget

Autonomous AI agents get worse at finding solutions as their own conversation piles up. That is the conclusion of Microsoft researchers, who call the effect "idea basins": local regions of the solution space that an agent with a long context can no longer escape. Their Fork-and-Flush method wipes the chat history but keeps the workspace. On the same compute budget, it delivers results up to 66% better.

How an agent gets stuck in a basin

An agent that works on a task for a long time fills its context, step by step, with hypotheses, failed attempts, and intermediate conclusions. All of this becomes the premise for the next step. The model keeps developing the direction it has already chosen, even when that direction leads to a dead end, and every new turn reinforces the trajectory. The researchers describe this as falling into a local region with no way out under the normal course of work.

A bigger budget does not solve the problem. If an agent keeps circling one idea, extra tokens and steps only deepen the basin. How the compute is spent therefore matters.

Fork-and-Flush: branching and resetting

The method consists of two operations. Fork creates branches of the process, and Flush erases the chat history. The key detail: only the conversation is erased, while the workspace stays untouched. Everything the agent has already built by hand survives: files, code, artifacts. The model gets a clean context and looks at the ready-made working environment without the baggage of its earlier reasoning.

The agent thus returns to the task without accumulated biases, yet loses no tangible progress. That is where the efficiency comes from: there is no need to restart the work from scratch.

Key parameters according to the study:

• Problem: "idea basins" caused by a growing chat context;
• What is reset: the agent's conversation history;
• What is kept: the workspace;
• Budget: the same amount of compute as in the baseline approach;
• Result: a gain of up to 66%.

"Up to 66%" means a maximum, not an average. The original publication in The Decoder gives no breakdown by task, model, or benchmark, so it is too early to judge the typical gain. You need to read the Microsoft paper itself.

What to try in your own agents

The idea applies well beyond a research lab, especially in coding, where agents work for hours. If the results hold up in practice, several steps follow from them.

First: store state in the workspace, not in the dialogue. Everything important (code, notes, test results) should live in files and the repository, where it can be read back after a context reset.

Second: do not stretch a single session indefinitely. If an agent makes no progress for a long time, it is cheaper to clear the context and give it a fresh look at the project's current state than to keep pouring tokens into the same dialogue.

Third: run several branches with a clean context instead of one long one. The compute budget stays the same, but you get more attempts at finding a different "basin."

The method challenges the common assumption that accumulated context always helps. For agentic systems, it is apparently better to treat context as a resource with diminishing returns, one to be managed as deliberately as tokens and run time.