Chinese AI company MemTensor has released and open-sourced MemT 1.0, a family of what it calls memory-native foundation models. The series includes two models: MemT-1.0-Flash, with 288 billion total parameters and 17 billion active, and MemT-1.0-Nano, with 36 billion total and 3 billion active. Weights are on Hugging Face and code is on GitHub alongside a technical report.
The models target a familiar cost problem in long conversations. Most systems carry the full history in the context window, so every additional turn makes inference slower and more expensive. MemT instead adds a trainable memory module inside the model. After each interaction, the model writes new information into a fixed-size memory state and recalls it when needed, rather than rereading the full record each time.
MemTensor says the models were trained in two stages on 34 billion tokens: first learning to write information into memory, then learning to answer questions from it.
All performance figures come from the company's own evaluation. MemTensor says end-to-end inference is about 1.86 times faster than a full-context approach at 128K tokens of history and about 3.99 times faster at 256K. It reports query time to first token holding at roughly 104 to 112 milliseconds, compared with about 445 milliseconds for full context at 256K. The company also says per-session memory state at 256K history is about 40 MiB, versus about 5.06 GiB for a full-context cache, a reduction of roughly 130 times.
On 11 memory benchmarks, MemTensor reports average scores of 50.6 for Flash and 47.0 for Nano, ranking first and second among the parametric memory methods it compared, and roughly level with retrieval-augmented approaches on average without retrieving history at query time. It says adding memory did not degrade Flash's general capabilities under distractor-heavy test conditions.
Developers should check the licensing details: the repository code is published under a PolyForm noncommercial license, while the model weights carry Apache and MIT licenses respectively.
The team says its next steps are improving long-term memory, expanding memory capacity and lowering the cost of memory updates, with the longer-term goal of agents that keep learning from their interactions.
SEE ALSO: Tencent Hunyuan Open-Sources Hy Memory 2.0 · METASTONE Launches TokenFleet