Event

Memory Tensor open-sources MemT 1.0 memory-native models

2 reports · 2 sources

Memory Tensor released and open-sourced the memory-native model MemT 1.0 series, comprising two models, MemT-1.0-Flash and MemT-1.0-Nano, with parameter scales of 288B and 36B and activated parameters of 17B and 3B, respectively. The series incorporates a persistent, updatable internal memory into model computation, allowing historical information and task experience to be automatically invoked in subsequent related tasks without the need to provide the original historical text again. According to the paper indexed by HuggingFace Daily Papers, MemT-1.0 is built on Metis and adopts the Gated DeltaNet-2 and Mixture-of-Storage architectures, with a training data scale of 34B tokens. The paper states that at a history length of 128K tokens, MemT-1.0-Nano achieves approximately 1.86x end-to-end inference speedup and 66x persistent storage reduction relative to full-context KV cache; evaluation results show that its average performance outperforms parametric memory baselines and is comparable to retrieval-augmented methods, while largely preserving general capabilities.