Chinese AI lab MemTensor has published MemT-1.0, a pair of language models designed to remember earlier conversations without carrying the full transcript along with every request. The technical report, MemT-1.0: Towards Memory-Native Evolution, was posted to arXiv on October 19, 2026 with 74 listed authors, and the company formally announced the open release on October 20. Code is on GitHub and the weights are on Hugging Face.
Today most assistants handle memory in one of two ways. They either keep stuffing old messages into the context window, which makes each call slower and more expensive as the history grows, or they bolt on an external store and retrieve snippets when needed. MemT-1.0 takes a third route. Each model carries a memory state of fixed size that is updated at the end of every turn, so earlier exchanges can be cleared from the prompt while their content stays available to the network.
Two sizes are available. MemT-1.0-Nano sits on the Qwen3.6-35B-A3B backbone, with 36 billion total parameters, 3 billion active per token, 451 million trainable memory parameters and a dynamic memory state that the README sizes at 10M. MemT-1.0-Flash uses DeepSeek-V4-Flash as its base and reaches 288 billion total parameters, 17 billion active, 3.61 billion memory parameters and a 336M dynamic state. In both, a Native Memory Block runs next to the backbone's attention. A Local Memory Block holds a multi-channel matrix that changes from step to step, a Hyper Memory Block supplies learned projections plus erase, write, forget and output gates, and updates are applied token by token with a stabilized form of Gated DeltaNet-2. A design MemTensor calls Mixture-of-Storage spreads that memory across several experts to raise capacity.
Training ran in two stages, continued pre-training followed by post-training, on 34 billion tokens in total. On 11 memory benchmarks, MemTensor says Flash and Nano finished first and second among the parametric-memory methods it compared, averaging 50.6 and 47.0 points respectively. The repository is candid about the ceiling: strong text-based memory systems, which retrieve the original evidence, remain ahead overall, especially on long histories where exact details matter.
The headline efficiency figures come from the report. At 128,000 tokens of history, Nano runs end to end about 1.86 times faster than a full-context baseline and needs roughly 66 times less persistent storage than keeping the KV cache. Figures circulated with the launch and reported by NetEase put the gain near 3.99 times at 256,000 tokens, with a per-session memory state of about 40 MiB.
MemTensor also tested the idea on coding agents. Instead of pasting lessons from earlier attempts into the prompt, its framework writes that procedural experience into memory when the agent gets stuck. The company reports that Nano's average success on SWE-bench Verified rose from 66.07% to 69.73% this way, slightly above a variant given the same experience as text, and from 60.00% to 63.56% on SWE-bench Multilingual.
Licensing is split. Flash weights carry an MIT license and Nano weights Apache 2.0, while the repository code, including the SGLang serving integration, is under the PolyForm Noncommercial License, so commercial use of that software needs separate permission. Every benchmark and efficiency number so far is self-reported, and no outside group has reproduced them.