An agent's memory is not only the rows you wrote. It is the summaries and extracted fields built from them, and deleting the row you addressed leaves its descendants in place. Against a matched agent that never saw the value at all, one plain question pulled a deleted secret back out of a 7B model in 68% of profiles, against 0 of 200 for the control — and when the summary held only three labelled pieces and the public rule for joining them, with no contiguous copy anywhere the model could see, exact reconstruction still ran 62–99%. Retiring every derived record descended from the deleted row took all of it to zero.