An arXiv preprint reports that MO-IKE balances fact edits and knowledge retention in frozen language models, with mixed results across benchmark datasets.