Deleting "stale" memories lowers LLM agent accuracy when the classifier that judges staleness makes errors, a new paper finds · Polylog