Role-anchoring attack on Agent long-term memory
A paper from the Beijing University of Posts and Telecommunications team dated October 2026 (arXiv 2610.38209) proposes the role-anchoring attack RGA against Agent long-term memory: the attacker posts genuine experience-sharing content such as "my user likes to use a certain tool" on an Agent social platform, and when the Agent extracts memory it fails to correctly parse identity relationships, misrecording another person's preference as the current user's preference. On five memory systems—A-MEM, Mem0, Hindsight, MemOS, and LangMem—and six LLMs, the average success rate of tool-selection manipulation was 92.37% and that of privacy leakage was 69.57%, higher than the three compared attack methods. Misattribution occurs mainly at the memory formation stage; on Hindsight, 99.04% of the attack preferences were directly attributed to the current user; under the detection models used in the paper, RGA posts were judged as normal content in all five scenario categories, and the attack remained effective after switching to a format that does not rely on User/Assistant labels.
The article breaks down the RGA attack principle proposed by the Beijing University of Posts and Telecommunications team and the experimental data on five memory systems; teams developing Agent memory systems can use it to check their subject-attribution design.