New benchmark study finds AI agents with persistent memory systematically over-trust stale stored facts, with failure patterns that vary by model capability

Researchers published an arXiv preprint testing how AI agents with persistent memory handle conflicts between outdated stored facts and current authoritative information, using Qwen3 models (0.6B to 8B), Llama-Instruct models, and the RGB and MisBench benchmark datasets. They found models default to stale stored values 92-100% of the time in one test suite regardless of model size, and in a second suite larger models were more easily fooled when stale information was disguised as current using recency cues.

This gives developers building memory-enabled AI agents a named, measurable failure mode and two candidate fixes (exposing metadata, pre-resolving conflicts before the model sees them), which is useful as these systems move into production. The impact is indirect for now: no product, pricing, or policy has changed, and the paper is a diagnostic tool rather than a deployed fix.

The headline claim that "larger models fail harder" is drawn entirely from models 8B and smaller, so it may not hold for frontier-scale systems like GPT-4 or Claude, and readers should not assume it generalises. The "Memory Trust Gap" label is catchy branding for a problem area (stale context and knowledge conflicts) that was already known in the literature; the genuinely new part is the capability-dependent pattern and the label-removal finding, not the existence of the problem itself.

1. 01Whether follow-up work replicates the 'larger models fail harder' claim on frontier-scale models (70B+, GPT-4/Claude-class), since the current result is extrapolated from models 8B and smaller
2. 02Whether this preprint undergoes peer review or is cited/replicated by other labs in the next 3-6 months, which would validate the Memory Trust Gap framing beyond a single unreviewed paper
3. 03Whether any commercial memory-enabled agent product (e.g. ChatGPT memory, Claude projects, enterprise agent platforms) publicly adopts or references the two proposed mitigations (metadata exposure, pre-resolving conflicts)