The problem: dirty memories and worlds that preserve the error
Agents based on language models that use external tools also accumulate their own mistakes in their history. When a system compacts, repairs, or passes this history to another model, it must decide how to recount those errors, while their effects remain in the environment. This is the starting point of the study published on arXiv under the title Cleaning an Agent's History Hurts When the World Still Holds the Error, of which the source reports only the abstract and metadata.
According to the authors, previous work assessed history cleaning in a fixed world—that is, under a condition in which showing a corrected step and declaring it completed are one and the same. The new question is whether the same correct content helps or harms depending on how the history presents it and on what the environment actually contains.
Two ways to correct and two opposite outcomes
The described method freezes an agent midway through a task together with the exact state of its world, modifies the history or the environment, and lets a receiving model continue. On AgentCE-Bench, after the agent has entered an incorrect value, the world is held fixed and the correct value is added in two natural ways: a truthful note next to the preserved error, or the action carried out by the agent itself.
The reported results are asymmetric. The note raises recovery—that is, the share of remaining goals achieved by the receiving model—from 51% with a neutral control note to 66%. The action marked as completed lowers it below the unmodified history, from 50% to 44%. For the authors, the mere presence of the correct value does not predict the effect: what matters is the form in which the history states it.
AppWorld: receipts and real effects
On AppWorld, after a database write, the experiment varies whether the receipt reports success or failure and whether the effect is actually present. A success receipt reduces full-task success from 45% to 28% when the effect is missing, while it increases it from 38% to 49% when the effect is present.
The difference between the two effects is 27 percentage points, aggregated across two receiving models. The source specifies that the gap is positive in six out of eight receiving models at 97.5% and in seven out of eight at 95%. These are measures stated by the authors in the same abstract and not replicated here.
What changes and what remains to be verified
The consequence indicated by the authors is that the effect of a modification to the history depends on the world that the history describes. A history made to appear clean while the environment still preserves the error can leave the agent in worse conditions than no modification at all. However, the work does not clarify which mechanisms generate the asymmetry, nor whether the results hold beyond the two benchmarks cited. Still to be verified are applicability to real systems and evaluation procedures—information that the abstract does not detail.
