A preprint study finds forget–retain update structure tracked language-model relearning better than weight distance, with utility and scope limits.