MRMS: A Multi-Resolution Memory Substrate for Long-Lived AI Agents

MRMS proposes a memory substrate for long-lived AI agents organized along two axes: representation, covering structured records, vectors, and graph relations; and time, covering short-term traces, medium-term abstractions, and long-term semantic commitments. The prototype combines these mechanisms for pre-generation memory selection, revision, boundary enforcement, and evidence attribution. Structured records determine eligibility, vectors support recall, and graph relations assess support, contradiction, and supersession before selected memory is projected into context. The abstract describes controlled long-lived interaction scenarios but reports no quantitative results.