EngramEdit Enables Knowledge Updates in LLMs via Decoupled Conditional Memory

Researchers introduced EngramEdit, a novel knowledge editing framework that updates factual knowledge in large language models by modifying decoupled n-gram memory embeddings instead of altering the main Transformer backbone. The method achieves near-perfect editing success while preserving unrelated model capabilities even as edits accumulate. Updating factual errors in traditional LLMs usually requires costly retraining or risks degrading model performance due to catastrophic forgetting. By turning conditional memory architectures into editable knowledge interfaces, EngramEdit enables efficient factual updates and improves multi-hop reasoning under chain-of-thought prompting by nearly three times compared to existing baselines. EngramEdit computes target memory representations for an updated fact across multiple phrasings and jointly updates the shared n-gram embeddings to match those targets. To protect unrelated facts, the algorithm heavily penalizes updates to frequently reused embeddings that are shared across different contexts.