An agent on a frozen model learns by writing notes into a file it re-reads on the next task. Take four published systems that do this, break their files into single claims, delete each claim and re-run the later tasks, and 12% to 33% of the entries turn out to have been making things worse. The fix in this paper never asks a model whether a claim is true. It admits a claim only when the environment has already backed it — which brings the harmful share down to 3.7% to 6.2% and raises the score on every model tested.