Memory Tensor open-sources MemT 1.0, a memory-native model that learns what to remember

On October 20, Memory Tensor released and open-sourced MemT 1.0, a memory-native large model that moves memory out of external retrieval modules and into the model's underlying architecture, so that accumulated experience becomes part of the model's own capability.

The team's definition of memory is deliberately narrow: storing chat logs is not memory. Its technical report uses a concrete case, where a user gives an exhibition's original time, later confirms it has moved to Sunday, and in between mentions an abandoned Monday alternative; the pair then talk for more than 500 turns, roughly 66,000 tokens, without the word "exhibition" coming up again.

External retrieval tends to mistake that passing remark for a conclusion. MemT 1.0, answering closed-book with no access to the original text, still reports both the original and rescheduled times, notes that the venue, ticket price and contact are unchanged, and keeps the final Sunday arrangement separate from the Monday alternative.

Architecturally the model splits its compute in two. A backbone network carries general world knowledge, the equivalent of factory settings; a dynamic memory state evolves in real time along a specific interaction trajectory, stays isolated between users and tasks, and takes part in every subsequent round of generation.

At the end of each turn the model refreshes that state through a token-level update mechanism. What to keep and what to overwrite is decided by gating mechanisms learned during training, and a parallel memory attention mechanism pulls the settled state back into the current computation when needed. No full retraining is involved.

Training took 34 billion tokens in two stages: 20 billion for continued pre-training, teaching the model to predict the future from fragments of the past, then 14 billion for post-training, where a stronger teacher model scored memory extraction and the errors were backpropagated. Storage uses a fixed volume, with a Mixture-of-Storage design in which a base MoE handles computation while multiple memory experts manage read and write allocation.

Across 11 system memory benchmarks covering long-term retention, dynamic updating and personalized understanding, with histories spanning 16K to millions of tokens, MemT-1.0-Flash and MemT-1.0-Nano averaged 50.6 and 47.0 closed-book, ahead of the strongest parameterized memory baseline in the same group by 12.5 and 8.9 points. Removing the post-training stage drops Nano from 47.0 to 36.3, and expanding storage experts from two to four adds up to 2.7 points.

Code and weights are published on GitHub and Hugging Face. For teams building long-running assistants and agents, the interesting question is not the benchmark gap but what changes once memory is internalized as a parameterized state, and whether closed-book memory holds up better than retrieval on real, long-horizon work.