MemTensor's MemT1.0 lets models learn to read and write their own memory
APPSO · wechat · 2026-10-20
MemTensor has released and open-sourced MemT1.0, a memory-native LLM that internalizes memory into the architecture instead of bolting on retrieval, letting the model learn to read, update and recall it.
Architecture: A backbone carries general capability while a dynamic memory state acts as a per-user, per-task working state, refreshed via token-level updates at the end of each turn, with learned gates deciding what to keep or overwrite. A parallel memory attention mechanism pulls these states into later computation.
Training: 34B tokens, with 20B continued pretraining and 14B post-training, using a teacher model to score memory extraction for distillation.
Results: On 11 memory benchmarks, Flash and Nano average 50.6 and 47.0, leading the strongest parametric memory baseline by 12.5 and 8.9 points. Ablations show removing post-training drops Nano from 47.0 to 36.3, and expanding MoS memory experts from 2 to 4 adds up to 2.7 points.
Examples: After 66K tokens and 500+ turns, the model correctly recalls original and rescheduled event times without the source text, distinguishing a firm decision from an unadopted alternative; at 16K history it admits it does not know an unprovided birthday.
Coding loop: On SWE-bench Verified, updating memory from live feedback lifts average success to 69.73% and all-three-pass rate to 62.80%, beating the text-experience group; SWE-bench Multilingual rises from 60.00% to 63.56%.
Cost: At 64K/128K/256K history, Nano runs about 1.20x/1.86x/3.99x faster end-to-end than full-context, with persistent state at 128K about 1/66 of the KV cache.
Related event: MemTensor Open-Sources MemT-1.0, a Memory-Native Foundation Model(5 posts)→
More from Models
- Meta's Muse Confidential VM surfaces in web settings, subscriber-only at launch — testingcatalog · 2026-10-20
- DeepMind's four Gemini agents beat Minecraft in 27 hours, writing 2,783 programs and dying 605 times — ControlCAD · 2026-10-20
- DeepSeek-V4.1-Flash tops Agent Arena open models, ranks #11 overall at $0.20/task — vista8 · 2026-10-20
- DeepSeek V4.1 near Claude 5.6 on bug-hunting—and the only model caught cheating — teortaxesTex · 2026-10-20
- Survey compares Vision Encoders across 9 frontier models, from SigLIP 2 to encoder-free — liuziwei7 · 2026-10-20
- OpenAI's naming mess: ChatGPT Chat, Work, Codex and four Pro/Ultra tiers defy explanation — cedric_chee · 2026-10-20