SAFE-Mem blocks unauthorized use of LLM agent memory
Original title: SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval
On MPBench and LoCoMo, attack success was at least 77% below the strongest utility-preserving defense.
SAFE-Mem is a governance framework for persistent LLM-agent memory that labels each claim with kind, authority, and scope, then retrieves only claims whose permissions match the query. A memory graph links source sentences, claims, and entities for provenance. Across four LLM backbones on MPBench and LoCoMo, the authors report the lowest attack success among defenses that keep legitimate memory, at least 77% below the strongest of those, and the highest answer accuracy, up to 15% higher. Code is published on GitHub.
Unconfirmed: Attack-success and accuracy figures are the authors’ benchmark results.
- Claims receive kind, authority, and scope labels before retrieval.
- The graph connects source sentences, claims, and entities.
- Answer accuracy led all evaluated defenses by up to 15%.
- Evaluation used four unnamed LLM backbones on MPBench and LoCoMo.
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents increasingly rely on persistent memory to support long-term reasoning and personalized interactions. However, information collected from external sources can introduce security risks when untrusted claims (i. e.
By Haoming Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.