MemTensor Open-Sources MemT 1.0, a Memory-Native Large Language Model

On October 20, 2026, MemTensor officially released and open-sourced MemT 1.0, a memory-native large language model. Unlike common retrieval-augmented approaches, MemT 1.0 internalizes memory into the underlying architecture, allowing the model to learn how to read, write, update, and recall memory through training.

The architecture splits the model into a backbone network and a dynamic memory state. The backbone handles general knowledge, while the dynamic memory state evolves independently across interaction trajectories, absorbing new information and participating in subsequent generation. MemT 1.0 refreshes memory through token-level updates, with a learned gating mechanism deciding which information to keep or update.

The team trained the model on 34 billion tokens in two stages: 20 billion tokens for continued pretraining and 14 billion tokens for post-training. They also introduced a Mixture-of-Storage (MoS) architecture, where multiple memory experts manage read/write allocation. On 11 systematic memory benchmarks, MemT-1.0-Flash and MemT-1.0-Nano scored 50.6 and 47.0 respectively, outperforming the strongest parametric memory baseline by 12.5 and 8.9 points. Ablation studies showed that removing post-training dropped Nano's average score from 47.0 to 36.3.

In programming scenario tests, MemT-1.0 Nano improved its average success rate on SWE-bench Verified from 66.07% to 69.73% by updating memory states based on real-time feedback. With a fixed-size dynamic memory state, MemT 1.0 is positioned as a foundation for long-horizon AI assistants and self-evolving agents.