Metis moves agent memory inside the backbone: gradient-free updates via a single forward pass, with frozen weights at inference

arXiv 2607.26760·low signalXBlueskyLinkedInCopy link

Agent memory today is almost entirely an external module — vector stores, file trees, retrieval layers bolted onto a frozen model. Metis (July 29, 2026) is a first prototype of a 'memory foundation model' that gives the backbone a persistent, dynamically evolving native memory state accessed through memory attention, with native store-and-use procedures acquired during mid-training. The operationally interesting property is that online memory maintenance is gradient-free — updates need only a forward pass and all learned weights stay frozen at inference — and the authors released checkpoints. Low importance for now because the abstract reports no benchmark numbers; this is a direction to watch, not a technique to adopt this week.