Skip to content
Book Open access

FortMem: Secure Stateful Data Management for Long-Term LLM Agent Memory

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 44 references

Abstract

Long-term LLM agent memory has become a critical state layer where records are written once and reused across sessions, turning prompt poisoning from a one-shot hazard into persistent database corruption. Traditional stateless RAG defenses reset per query and fail to evaluate the trustworthiness of persistent records over time. To address this, we present FortMem, a structurally robust memory layer, and PASTP (Provenance-Anchored Signed Trust Propagation), a zero-trainable-parameter trust score that formulates provenance as the teleport vector over a signed ANN--NLI corroboration graph. Theoretically, our Closed-Coalition and Open-Boundary Trust Caps prove that adversarial coalitions can inflate trust solely via their boundary fraction, a runtime-measurable structural channel. To counter advanced provenance spoofing, an enhanced variant (PASTP+) automatically audits and demotes the priors of spoofed records contradicted by trusted neighbors. Extensive evaluations across three real-passage memory stores demonstrate that PASTP maintains resilient retrieval correctness under persistent poisoning, whereas vanilla dense retrieval degrades catastrophically. Furthermore, our boundary diagnostic accurately detects 96.3% of poisoned snapshots with zero false alarms, and downstream evaluations confirm that FortMem severely limits poisoned-answer adoption without compromising standard generative capabilities.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.