Forgetting is not always bad: a neuroscience-inspired memory repair mechanism for poisoned LLM agents

The researchers propose a training-free memory repair framework called DREAM to repair the memory of poisoned LLM agents, reducing the attack success rate of backdoor poisoning by more than 95% under various attackers while maintaining benign task utility. DREAM is inspired by the mechanism in neuroscience whereby weak reactivation can destabilize memory and induce decay in the absence of reinforcement. It includes perturbation-induced implicit reactivation, adaptive anomaly diagnosis based on multidimensional activation patterns, and a dynamic repair module that selectively suppresses harmful memories and reinforces benign ones. The paper states that DREAM also has some robustness to injection attacks. Compared with A-MemGuard, it achieves up to 2.13x speedup and 37.7% lower token consumption, and reaches a task success rate of 92.96% on poisoned multi-agent systems.