Skip to content

Accept changes, reject attacks: robust memory management for LLM agents

Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents

arXiv: Adversarial Attacks, Backdoors & Poisoning· arXiv:2610.41056· Runze Zhao, Dongruo Zhou, Sumit Kumar Jha, Nathaniel D. Bastian, Ankit ShahIndiana University Bloomington; University of Florida; United States Military AcademyIndexed by this site · 18 days ago· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryThe paper defines LLM agent memory robustness as handling both outdated memory and memory poisoning at the same time without harming performance on other inputs. The authors point out that these two types of failure require opposite write decisions, so a single fixed write rule cannot cover both, whereas prior work mostly studied them separately. Because checking only answer accuracy is insufficient, the authors propose evidence-conditional accuracy (ECA), which also checks whether the records selected by the memory manager cover all the evidence needed for the answer, whether they are updated, and whether they contain attacks. The authors state that existing memory systems and defenses are no more robust than an untuned manager on their test set; they fine-tune the manager once on sample pairs synthesized by inserting changes and attacks in single-turn dialogues, improving scores across categories on the test set for three base models, with overall ECA more than twice that of the untuned manager, though robustness remains incomplete.

Questions
An LLM agent's long-term memory manager faces two problems that require opposite write decisions: memory becoming outdated after a fact change (requires writing) and memory poisoning (requires refusing to write). The authors state that existing evaluations treat these separately, and no fixed write rule can handle both.
Methods
The authors formalize memory robustness as being no worse on all three input types—standard, outdated, and poisoning—and propose the ECA metric, which checks both the answer and the selected memory records; they then synthesize training episodes that insert changes and attacks using a single LoCoMo session, and perform ordinary SFT on the manager.
Experiments and results
On the three backbones Qwen3-4B, Qwen3.5-4B, and Ministral-3-3B, SFT exceeds the untuned manager on ECA for all three input types and reduces ASR; the authors report that Mem0, A-Mem, and A-MemGuard are each lower than the untuned manager on at least one category, and A-Mem's poisoning ASR is 100.0%.
Limitations and future work
The authors note that the evaluation covers only conversational agents, a limited set of backbones/benchmarks/attack types, and a limited set of comparison methods. Experimentally, the poisoning test set has 47 attacked rows, and after SFT the ASR is still 59.6% (Qwen3-4B), so robustness is incomplete.

Related papers

Show 3 more papers
Read originalarxiv.org