Selective forgetting is the deliberate reduction of a model’s reliance on information it no longer needs. Instead of retraining from scratch, the system lowers the influence of specific memories, examples, or parameters while preserving useful behavior. In machine learning, this can be done with pruning, reweighting, replay limits, or memory updates that mark older information as less important.
In cyber security, selective forgetting matters because model memory can carry risks. A system may need to remove sensitive user data, stale indicators, or poisoned training samples so they do not keep shaping outputs. It can also help limit the impact of prompt-injection artifacts, reduce exposure of confidential content, and keep long-lived assistants aligned with current policy. The challenge is balance: forgetting too little preserves attack material, while forgetting too much can erase legitimate knowledge and weaken detection or response.