Metis Puts Memory Inside the Weights: a 27B 'Memory Foundation Model' With Dynamic Parameter Updates, Code and Checkpoints Released

arXiv / HuggingFace Daily Papers (211 upvotes)·medium signalXBlueskyLinkedInCopy link

A MemTensor Shanghai / Renmin University / NUS / SJTU / Tongji team posted Metis (arXiv:2607.26760) on 2026-07-29, framing it as the first memory foundation model — memory as a native computation rather than a bolted-on retrieval module. Metis blocks pair local memory (a memory matrix M with normalization vector S) with hyper-memory blocks that aggregate hidden states into memory key-value pairs, consumed by attention that blends standard and memory attention, with parameters updating as persistent state across interaction steps. Metis-27B reports 24.76% on MemOps no-context, 26.74% on LoCoMo Gold, 50.82% on NextMem and 18.56% on out-of-distribution ATM-Bench; code and checkpoints are public under CC BY-NC-SA 4.0 at github.com/MemTensor/Metis.