“The "memorization sinks" technique trains a model by designating specific neurons that are updated only on a particular document, isolating that document's information to enable easier unlearning.”