MemTensor's Metis debuts native-memory foundation model

An arXiv paper out this week from a Chinese startup called MemTensor and a group of academic collaborators makes a specific architectural bet about where agent memory should live. Instead of the now-standard pattern of pairing a frozen language model with an external vector store, the paper argues for what the authors call a memory foundation model: a persistent, dynamically evolving memory state carried inside the backbone itself, with maintenance done through forward passes while the weights stay frozen.

The system they build to demonstrate this is called Metis. Per the paper, it is evaluated at 4B, 9B, and 27B parameters on a Qwen3.5 backbone, trained on 8 H100 GPUs against a synthesized dataset of roughly 406M tokens covering four memory operations the authors name explicitly: remember, forget, update, and reflect. The headline number is that Metis-27B reports 73.77% average on the authors' constructed test set, against 1.69% for the same-size Qwen3.5 baseline without context. The paper also lists evaluations across MemOps, LoCoMo (Gold), NextMem, SQuAD, HotpotQA, LongMemEval and LongBench, and the authors commit to releasing project materials and model checkpoints.

Why this is worth watching even if you do not build agents: most of the last two years of agent-memory progress has been retrieval plumbing wrapped around frozen models. If native memory as an architectural component actually holds up, the memory layer stops being a database problem and starts being a model-design problem, and the edge shifts from infrastructure vendors toward model builders. MemTensor, with academic co-authors from Renmin University of China, NUS, Shanghai Jiao Tong University and Tongji, is trying to plant a flag there early.

The honest caveat is that the standout gap is on the authors' own constructed test set, and a self-described first step plus a small compute footprint means the harder comparisons are not settled here. What the paper does not give you is a head-to-head against strong RAG stacks on the standard long-memory benchmarks it names, the inference-cost profile at 27B, or license terms and dates for the promised checkpoints. Take the specifics as reported, not as validated.

The part worth watching is not the leaderboard number. It is whether other groups pick up the memory-inside-the-backbone framing at all, because if they do, the 'vector DB plus a frozen LLM' pattern starts to look like scaffolding rather than the destination.