ModelScope releases MemT-1.0, open-weight models with built-in persistent memory

ModelScope announces MemT-1.0, an open-weight model family that keeps persistent, evolving memory inside the model's computation instead of reloading the original context.

Conversations and agents can reuse historical information this way, according to the announcement on X.

The post says MemT-1.0-Flash and MemT-1.0-Nano rank first and second among evaluated parametric memory methods across 11 memory benchmarks. The post does not name the benchmarks or the other methods compared.

Flash weights are released under the MIT license, and inherited backbone licenses still apply.