Long-horizon agents accumulate records of their own failures as they work toward distant goals. A new arXiv preprint, "When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents," argues that the usefulness of any particular failure record is not fixed: it can shift as the agent gets further along in a task. The authors frame this as a decision problem: for a given block of history, should the agent keep it as-is, compress it, or remove it entirely?
According to the abstract, the work compares how each of those choices affects subsequent execution. The key move is treating forgetting as an adaptation step rather than a simple memory limit. Because the value of a failure record may decline or change in meaning over time, the same block of history could help early on and hurt later.
At present the public abstract cuts off before describing the proposed method or results, so the full details of the risk control and evaluation are not available from this source. The central question, however, is clear: context should be managed dynamically, not stored indefinitely.