MemT 1.0
MemTensor open-sourced MemT 1.0 Flash and Nano, models with a trainable memory state that keeps long histories in fixed size, with about 4x faster inference at 256K history.
Most systems carry the full history in the context window, so each turn costs more. MemT adds a trainable memory module inside the model that writes new information into a fixed-size state after each interaction and recalls it later. Training used two stages on 34 billion tokens, first writing to memory, then answering from it.
- Date
- Thursday 22 October 2026
- Lab
- MemTensor
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Flash parameters | 288B total, 17B active MemT-1.0-Flash | company |
| Nano parameters | 36B total, 3B active MemT-1.0-Nano | company |
| End-to-end inference speedup vs full context | 1.86x at 128K, 3.99x at 256K history length in tokens | company |
| Per-session memory state at 256K | about 40 MiB vs 5.06 GiB full-context cache roughly 130x smaller | company |
| Average score on 11 memory benchmarks | 50.6 Flash, 47.0 Nano first and second among parametric memory methods compared | company |
| Query time to first token | about 104 to 112 ms vs about 445 ms full context at 256K | company |
Weights are on Hugging Face under Apache and MIT licenses respectively, while the repository code uses a PolyForm noncommercial license. All figures are from the company's own evaluation.
Sources
This record was checked against its sources on 22 October 2026. How we check