“In experiments with "memorization sinks," the technique was applied only within the MLP layers of the transformer architecture, based on the theory that factual information is primarily stored there.”