An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
MDL adds a white-box trust decision between retrieval and generation, using relevance, reliability, and task risk to reject conflicting memories without another model call.
The Memory Decision Layer is a **zero-parameter** controller between retrieval and generation. It combines relevance, reliability, and task risk, separates confidence from consistency, and can abstain instead of injecting a retrieved memory.
Agent builders should treat retrieval and trust as separate decisions, especially when stored memories conflict. In the reported evaluations, MDL reduced hallucinations by **about 56.04%** in general conflict scenarios and added **about 0.14 ms** per decision.
The Memory Decision Layer is a **zero-parameter** controller between retrieval and generation. It combines relevance, reliability, and task risk, separates confidence from consistency, and can abstain instead of injecting a retrieved memory. Agent builders should treat retrieval and trust as separate decisions, especially when stored memories conflict. In the reported evaluations, MDL reduced hallucinations by **about 56.04%** in general conflict scenarios and added **about 0.14 ms** per decision. The results span multiple models and open datasets, but the abstract does not identify them or describe production workloads. Near-zero hallucination in high-risk scenarios therefore remains a paper result rather than a deployment guarantee.
MDL makes memory admission a distinct, inspectable control after retrieval: relevance alone does not justify injecting a memory when reliability is low or task risk is high. It complements query-scoped suppression and provenance-aware stores with a cheap abstention point, but does not address model-upgrade portability or establish that its reported conflict results transfer to production workloads.