Large language model (LLM) agents are highly vulnerable to memory‑poisoning attacks. Existing defenses mainly rely on external modules for static memory isolation or continuous online auditing, which leads to low memory efficiency and high computational overhead. Inspired by neuroscientific findings that weak reactivation can destabilize memories and cause decay without reinforcement, we propose DREAM (Dynamic Reactivation for Engram Attenuation in Memory), a training‑free, agent‑intrinsic memory repair framework that enables selective functional forgetting of poisoned memories.
DREAM follows a three‑stage pipeline: (1) perturbation‑induced implicit reactivation to expose the structural fragility of malicious memories; (2) adaptive anomaly diagnosis based on multi‑dimensional activation patterns to detect topological anomalies; (3) a dynamic memory repair module that suppresses harmful memories while reinforcing benign ones.
Extensive experiments across diverse attackers and real‑world tasks show that DREAM reduces backdoor attack success rates by over 95% while preserving strong benign utility. It also achieves competitive robustness against injection attacks. Compared with A‑MemGuard, DREAM attains up to $$2.13\times$$ speedup and a 37.7% reduction in token consumption. In poisoned multi‑agent systems, the task success rate reaches 92.96%.
Review: DREAM leverages internal memory dynamics to provide an efficient, low‑overhead defense, highlighting the promise of neuro‑inspired mechanisms for LLM security.