A defense without re-training weakens malicious memories while consolidating useful ones. The results cover back door attacks and injections.
DREAM attempts to process an agent's memory directly rather than isolating it or having an external module audit each access. Small disturbances reactivate memories, a diagnosis identifies anomalies and a repair phase reduces the influence of suspicious elements.
The authors report a reduction of more than 95% in the success of backdoor poisoning attacks in their trials. Against A-MemGuard, they measure up to 2.13 times less execution time and 37.7% less tokens saved. These gains remain dependent on the attackers and tasks selected for evaluation.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →