Memory-based editing is a technique for updating or correcting factual knowledge in language models by storing modified facts in an external memory rather than altering the underlying model parameters. During inference, the system retrieves relevant updated facts from this external memory and supplies them to the model, typically via in-context retrieval or iterative prompting, enabling the generation of answers that are consistent with the new information. By decoupling knowledge updates from the internal weights, this approach avoids catastrophic forgetting and parameter corruption, allowing language models to accurately reflect new or changed information while maintaining their general reasoning and generation capabilities.
