Study Finds Editing AI Agent History Can Lower Task Success by 27 Points When World State Still Contains Error

Researchers tested tool-using language-model agents by freezing them mid-task with exact world states, then editing either the agent’s history or the external world and letting a receiver model continue. On AgentCE-Bench, after an agent committed a wrong value, they added the correct value in two ways: a truthful note beside the preserved error, or an edit showing the agent’s own completed action. On AppWorld, they varied database-write receipts and whether the write’s effect was present.

The paper reports that a truthful note raised recovery, the share of remaining targets a receiver got right, from 51% under a matched neutral note to 66%. The completed-action edit lowered recovery below the unedited history, from 50% to 44%. In AppWorld, a success receipt lowered whole-task success from 45% to 28% when the database effect was absent, but raised it from 38% to 49% once present. The pooled difference was 27 points across two receivers.

The result challenges a common assumption in agent engineering: that removing or repairing an error in the transcript is automatically beneficial. Earlier work cleaned histories at a fixed world, where corrected steps and world updates moved together. This study separates those changes. It shows the same correct value can help or hurt depending on how the history states it and whether the environment already reflects it. The effect appeared in six of eight receivers at 97.5% and seven at 95%.

For teams building long-running autonomous systems, the findings indicate that memory hygiene cannot be treated as simple deletion or cosmetic correction. Developers face a coordination problem: history edits must align with world-state checks, especially after tool calls, database writes, or handoffs between models. A cleaner-looking transcript can mislead a successor model if the external environment still contains the mistake, leaving the agent worse off than if the history had been left unchanged.