VeriMem 2.0: Verification-Gated Declarative Memory for LLM Agents — Prevention at Write, Not Detection at Use
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
LLM agents persist declarative memory—facts, rules, preferences—across episodes, but existing write policies admit entries on the model's own judgment (LLM-judge scoring; success filters), and both write hallucinations: under matched extractors and retrievers, ungated writes contaminate 12.8% and LLM-judged writes 17.9% of audited location claims on ScienceWorld (pooled over three seeds), while the best LLM agent answers memory-validity probes correctly only 55.2% of the time. We present VeriMem 2.0, a memory admission gate that requires each declarative write to carry an executable evidence predicate—a decidable check over a logged evidence corpus—and to abstain when no predicate is expressible. In a controlled comparison where only the gate varies, VeriMem reduces contamination to 1.5% on ScienceWorld (0.0% on ALFWorld) and sustains this under adversarial injection of up to 80 forged entries (dose 0.4), landing zero of them on both environments while ungated and success-filtered stores absorb 100%. It holds success rate under attack while ungated memory degrades on ALFWorld (0.35 vs. 0.15). Ablations attribute the entire defense to the executable predicate: removing it collapses VeriMem to no-gate and admits 32/40 forgeries, while removing evidence-log retraction changes nothing (0/40). The gate costs clean success rate within noise (0.140 vs. 0.113 ScienceWorld; 0.22 vs. 0.28 ALFWorld) and buys a 3–5× smaller, 100%-verified store; its abstention (35.2% ALFWorld, 32.1% ScienceWorld) is a tunable fallback ablated directly. All experiments use a 7–8B open-source backbone (Qwen3-8B). Code, prompts, and audit tooling will be released.
- Primary area:
- foundation or frontier models, including LLMs
- Keywords:
- LLM agents, declarative memory, verification, hallucination, memory governance