Neural-network updates for new data can overwrite parameter patterns that supported older skills. The model may improve on the latest task while performance on earlier tasks drops sharply.
Mitigations include replaying earlier examples, protecting important parameters, modular architectures and regular evaluation across old and new tasks. No method is useful unless it measures retention over time.
ELI5
Catastrophic forgetting is when an AI learns something new but loses an important skill it had before. The new training changes shared model values that the older skill depended on.
For example, an assistant trained on a new document format might suddenly become worse at the original format. Teams can replay older examples and test both skill sets after each update to catch the loss.
