Label agent memory epistemically — separate what the user told you from what you retrieved (MRMS)

arXiv 2607.04617·low signal

MRMS (arXiv:2607.04617, July 6, 2026) organizes long-lived agent memory on two axes at once: representational (structured records, vectors, graphs) and temporal (short-term traces, medium-term abstractions, long-term semantic commitments). The design worth stealing is the division of labor — structured records govern eligibility, vectors handle recall, and graphs resolve contradictions and supersessions *before* anything is projected into context. Its fourth requirement is the one most memory systems skip: epistemic labeling that distinguishes personal context from external evidence, so an agent never launders something it read on the web into something the user said. This is an 11-page technical report with a prototype, not a benchmarked system — treat it as an architecture reference.