MemTensor releases MemT 1.0 models with native memory branches inside Transformer modules

What happened

MemTensor has open-sourced MemT 1.0, which adds trainable memory branches to selected Transformer modules to encode, update, and retrieve experience as internal dynamic states. MemT-1.0-Nano has 36B total parameters with 3B active, while MemT-1.0-Flash has 288B with 17B active. Across 500 SWE-bench Verified tasks run three times each, Nano averaged 69.73% success versus 66.07% without experience and 69.13% with text experience; the article acknowledges that text memory systems retain advantages in recalling details from long histories.

The story so far

MemTensor has open-sourced its MemT 1.0 model series, adding trainable memory branches to selected Transformer modules so the models can encode, update and retrieve experience as internal dynamic states. The series includes MemT-1.0-Nano, with 36B total parameters and 3B active, and MemT-1.0-Flash, with 288B total parameters and 17B active. QuantumBit reported that, across 500 SWE-bench Verified tasks run three times each, Nano averaged a 69.73% success rate using internal experience, compared with 66.07% without experience and 69.13% with text experience. The article also acknowledged that text memory systems retain advantages in recalling details from long histories.

Keep reading

Trending now

1. 1Mistral Releases Voxtral 3 Speech Models, Voice Agents API and LiveKit-Based SDK
2. 2DeepSeek Open-Sources DeepInfer and API Components, With Model and Hardware Limits
3. 3OpenAI Acquires Orchestra Bio to Expand AI Tools for Drug Development
4. 4Google Rolls Out Gemini Personal Intelligence to Android Auto, With Google Built-In Cars to Follow