MemT-1.0 builds memory into the model itself
A new arXiv paper introduces a family of memory foundation models that treat memory as part of model computation, with released Nano, Flash and Pro variants.
In plain words
Most AI systems handle memory by storing past information outside the model and retrieving it when needed. The MemT-1.0 paper instead describes models that accumulate and use information internally as part of their own computation. The authors report this approach is faster and needs less storage than keeping full conversation context.
What was announced
The paper, published on arXiv, defines memory-native evolution as making memory an intrinsic capability of foundation models so they can continuously accumulate and use information. MemT-1.0 builds on Metis and uses Gated DeltaNet-2 with a token-level memory update mechanism inside each turn, while a Mixture-of-Storage component scales memory capacity. Training followed a two-stage pipeline of continued pre-training and post-training, supported by 34B tokens of data. The released family comprises MemT-1.0-Nano, MemT-1.0-Flash and MemT-1.0-Pro.
In the authors' evaluation, the models outperformed parametric memory baselines on average and stayed competitive with retrieval-augmented methods while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano showed roughly 1.86x end-to-end inference speedup and a 66x reduction in persistent storage compared with a full context KV cache. The authors also report improved task success on agentic tasks when retrieved procedural experience is internalized.
What it does not show
- The source is the paper's abstract only, so evaluation methods, baselines and model availability are not described in the supplied evidence.
- The reported speedup, storage and performance figures are the authors' own claims and have not been independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.