An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
The paper proposes a no-training memory gate that decides when retrieved agent memories are too risky to use.
Its Memory Decision Layer sits between retrieval and generation, scoring retrieved memories for relevance, reliability, and task risk. The authors argue that ordinary RAG can worsen hallucinations when the memory store contains conflicts. In their tests, MDL cut hallucinations under conflicting memories by about 56.04% and neared zero in high-risk settings. It is described as white-box, geometric, and fast, adding about 0.14 ms per decision. ArXiv · AI/CL/LG's note
Its Memory Decision Layer sits between retrieval and generation, scoring retrieved memories for relevance, reliability, and task risk. The authors argue that ordinary RAG can worsen hallucinations when the memory store contains conflicts. In their tests, MDL cut hallucinations under conflicting memories by about 56.04% and neared zero in high-risk settings. It is described as white-box, geometric, and fast, adding about 0.14 ms per decision. ArXiv · AI/CL/LG's note
score 5