- Published on
Betting on Memory-Native: MemT 1.0 Breaks Through in the Second Half of Growth-Oriented AI
- Authors
- Name
- DeepHorus
Memory Tensor officially released and open-sourced the memory-native large model MemT 1.0, and simultaneously launched MemT-1.0-Flash and MemT-1.0-Nano. The model internalizes memory into the underlying architecture and learns through training to read, write, update, and invoke memory, attempting to solve the problems of AI repeating mistakes across sessions and failing to accumulate experience.
Against the backdrop of Meta Muse, OpenAI Dots, and xAI Grok Bot setting off a wave of private assistants, mainstream external retrieval is said to be able only to search through static memos and to find it difficult to consolidate the tuning process into the working tacit understanding required for a long-term assistant. MemT 1.0 chose the memory-native route, making experience accumulation part of the model's capabilities.
Architecturally, MemT 1.0 splits its capabilities into a backbone network and a dynamic memory state. The dynamic memory state is independent across different users or tasks, absorbs information during continuous interaction, and directly participates in subsequent generation computation. At the end of each dialogue round, the model refreshes its memory state through a token-level update mechanism, with a gating mechanism learned through training deciding whether to retain or modify it.
The technical report disclosed a test: the user first gave the exhibition's originally scheduled time, later finalized a rescheduling to Sunday, and mentioned a Monday alternative that was not adopted. After more than 500 rounds and about 66,000 tokens of chat, MemT 1.0, in a closed-book state detached from the original text, still accurately reported the originally scheduled and rescheduled times, and distinguished the Sunday final decision from the Monday alternative.
The team trained it with 34 billion tokens in two steps and introduced the Mixture-of-Storage architecture. On 11 systematic memory benchmarks, Flash and Nano achieved closed-book average scores of 50.6 and 47.0 respectively, leading the strongest parametric memory baseline in the same group by 12.5 and 8.9 points. Ablation showed that after removing post-training, Nano's average score dropped from 47.0 to 36.3.
In the 500 tasks of SWE-bench Verified, pre-writing experience raised Nano's average success rate from 66.07% to 69.33%, and its all-three success rate from 54.00% to 61.60%. In real-world use, after updating the memory state based on real-time feedback, the average success rate was 69.73%, and the all-three success rate was 62.80%. In the 300 tasks of SWE-bench Multilingual, the average success rate rose from 60.00% to 63.56%, and the all-three success rate jumped from 45.33% to 55.33%.
In terms of engineering cost, when processing 64K, 128K, and 256K token histories, Nano's end-to-end inference speed was approximately 1.20x, 1.86x, and 3.99x that of the full-context solution, respectively. At 128K, persistent state storage was about one sixty-sixth of the full-context KV cache. Memory Tensor believes this provides a foundation for long-cycle applications such as enterprise R&D experience reuse, digital avatars, chronic disease management, and personalized private tutoring.
