DREAM Framework for Selective Forgetting of Poisoned LLM Memories

DREAM provides a training-free, agent-intrinsic mechanism to repair memory-poisoning attacks in LLM agents. It uses perturbation-induced reactivation and anomaly diagnosis to selectively decay malicious memory engrams without the overhead of external auditing modules.