An agent can turn a mistaken lesson into a permanent instruction. That instruction can shape later executions, whose saved records then regenerate the same mistake. Ruhan Wang, Kishan Panaganti and Dongruo Zhou study this possible feedback loop in agent memory.
Their Anti-Persistence design saves past actions and outcomes. For each new task, a selector chooses direct use of those records or temporary advice derived from successes, strategies or failures. The advice is discarded after execution; the records and mode-selection feedback remain.
Across three models and three text-based task collections, the authors report:
- Average held-out success of 59.9%, compared with 53.5% for saved execution records, 49.0% for saved lessons and 49.3% for both.
- The best result in seven of nine model-and-task combinations, after ten memory-update rounds with model weights fixed. Evaluation freezes memory and all related state; results average three seeds.
- The lowest average task-time model-call count among those four memory designs, including advice generation. This measures requests, rather than billed tokens or elapsed time.
A direct comparison also favors discarding generated advice over saving it in four tested settings. The size of that benefit varies considerably across models and tasks.
The paper’s stronger conceptual claim needs care. Its formal analysis separates exposure, generated errors and admission to storage; it does not prove inevitable degradation. A separate audit of parent–descendant associations lacks verified event-level logs and cannot establish causal error reproduction.
Caveat: saved execution records can still carry mistakes into future advice. Temporary lessons remove one route for persistence, while retrieval quality and the value of externally verified reusable knowledge remain open questions.