MemTensor open-sources MemT-1.0, memory-native models built on DeepSeek-V4-Flash and Qwen3.6 that keep a fixed-size memory inside the model instead of resending history; text-memory systems still lead overall

Why it matters

【Key points】

Shanghai-based MemTensor (the MemOS team) released and open-sourced MemT-1.0, a "memory-native" model family, on Oct 20 under the MIT license.

1. Idea: a Native Memory Block sits beside the backbone's attention, with a dynamic local memory and a trainable hyper memory. Memory updates at the end of each turn, so later inference reads memory instead of the raw history. The state has a fixed size; MemTensor says it needs about 66x less storage than a full-context KV cache at 128K history.

2. Models: MemT-1.0-Nano on Qwen3.6-35B-A3B (36B total, 3B active, 451M trainable memory parameters); MemT-1.0-Flash on DeepSeek-V4-Flash (288B total, 17B active, 3.61B memory parameters, 336M dynamic state).

3. Two-stage training on 34B tokens.

4. Self-reported: across 11 memory benchmarks, Flash and Nano rank first and second among the parametric-memory methods tested. The model card says strong text-memory systems still lead overall, especially on long histories that need detailed evidence.

5. A demo stores retrieved experience in native memory for long-horizon coding agents. The tech report PDF ships with the weights; arXiv is "coming soon".

【Verifiable】

Flash: https://huggingface.co/MemTensor/MemT-1.0-Flash

Nano: https://huggingface.co/MemTensor/MemT-1.0-Nano

Code: https://github.com/MemTensor/MemT-1.0

【Our read】

1. An open attempt to move memory from bolt-on RAG into the model itself, built on open backbones others can reproduce.

2. By its own account text memory still wins on accuracy; the gain today is storage and not resending history.

3. All results are self-reported and unreplicated; Flash, at 288B total parameters, is heavy to self-host.