Bigger context windows are not the same as better memory — and a new research paper makes that case with an actual alternative.
Researchers have published a framework called MRMS, short for Multi-Resolution Memory Substrate, that organizes agent memory along two axes. The first is representational: structured records, vector embeddings, and graph relations working together. The second is temporal: short-term traces, medium-term abstractions, and long-term semantic commitments. The design forces those layers to stay in sync — structured records control what is even eligible for recall, vectors handle retrieval, and the graph layer resolves conflicts and flags when older memories have been superseded. The team also built a lightweight prototype that exercises all of this under controlled long-lived interaction scenarios.
The paper's core claim is that reliable personalization is a memory design problem, not a compute problem. That reframe matters because most current approaches either dump everything into a long context window — expensive and imprecise — or rely on vector retrieval alone, which can surface outdated or contradicted information without flagging it. A graph layer that tracks supersession is the part that is genuinely underexplored in production agent stacks.
Most commercial agent frameworks today treat memory as an afterthought bolted onto a chat log. If this architecture holds up outside controlled conditions, it is closer to how a useful assistant actually needs to work — which is also why it will take a while before anyone ships it.