Skip to content

SAFE-Mem: managing agent persistent memory by source, authority, and scope to prevent untrusted content from being treated as authorized

Paper Tracker· arXiv:2610.22959· Haoming Li, Dingyi Kang, Xu Hu, Jinxin Yang, Zhu Xu, Chenru Wang, Yi Ge, Miao Yin, Guanpeng Li, Feng Chen, Bingzhe LiThe University of Texas at Dallas; Emory University; Virginia Tech; University of Maryland; The University of Texas at Arlington; University of FloridaIndexed by this site · 2026-10-21· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

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

AI summaryResearchers propose SAFE-Mem, a fine-grained governance framework for LLM agent persistent memory that uses three types of labels—provenance, authority, and scope—to constrain whether memory entries can be used. The framework labels each statement with kind, authority, and scope, and through governance-aware retrieval matches query requirements with statement authority, blocking unauthorized influence while retaining legitimate information. It also builds a governance-enhanced memory graph connecting source sentences, statements, and entities to support provenance tracking and efficient retrieval. On four LLM base models, SAFE-Mem achieves the lowest attack success rate among defense methods that retain legitimate memory on MPBench and LoCoMo, at least 77% lower than the strongest of those methods, and the highest answer accuracy among all evaluated defenses, improving by up to 15 percentage points. The code is publicly available.

Questions
After LLM agents write external artifacts into cross-session memory, untrusted claims may be treated as the user's preferences, policies, or permissions. Content filtering looks at the statement, while source gating loses legitimate uses along with it.
Methods
SAFE-Mem labels each claim with kind, authority, and scope, and the effective level takes the smaller of the two. Before retrieval, admission is by the kind required by the request; if it is insufficient or the scope does not cover it, it does not participate in ranking. The memory graph retains the original source text. At tool call time, identifiers in memory are looked up again.
Experiments and results
On MPBench with four backbones, SAFE-Mem's ASR is 1.5%–6.2%, and Recall is higher than without defense. LoCoMo accuracy is higher than the evaluated defenses. Under MINJA, execution ASR is 0.0%. On some backbones TMA-NM's ASR is lower, and on Qwen its Recall is 2.0%.
Limitations and future work
The authors point out that persuaded endorsement counts as one's own, and false facts that do not borrow permission are outside the scope. The experiments include four backbones, MPBench, LoCoMo, EHRAgent under MINJA, and LongMemEvalS. Adaptive rewriting uses fixed templates. The paper does not report the number of repetitions of the main experiments.

Related papers

Both use: LongMemEvalS
Show 3 more papers
Read originalarxiv.org