Study Finds Editing an Agent's Error History Can Backfire If the World Still Holds the Mistake

A new arXiv preprint tests how compacting, repairing, or handing off a tool-using language-model agent's history affects its later performance. On AgentCE-Bench, a truthful note with the correct value raised recovery from 51% to 66%, but a completed action that reflected the correct value lowered it to 44%. On AppWorld, a success receipt for a database write lowered task success from 45% to 28% when the write's effect was absent, and raised it to 49% when the effect was present.

Why it matters: Developers who clean, summarize, or hand off agent histories should check that the edited history matches the actual state of the world, since a history that looks clean but hides a live error can leave the agent worse off than no edit.