MemTensor released and open-sourced MemT 1.0, a memory-native large language model that internalizes memory into its architecture, trained on 34 billion tokens with a Mixture-of-Storage design and top scores on 11 memory benchmarks.