MemT-1.0 memory foundation models: 1.86x faster, 66x less storage at 128K context
pmttyji · reddit · 2026-10-21
MemTensor released MemT-1.0, a family of memory foundation models (Nano and Flash) that internalize memory as part of model computation, built on Gated DeltaNet-2 with token-level memory updates and Mixture-of-Storage, trained on 34B tokens. They beat parametric memory baselines and match retrieval-augmented methods; at 128K history, Nano delivers 1.86x end-to-end inference speedup and 66x less persistent storage vs full-context KV cache, while improving agentic task success. Paper, code, and weights are open.
Related event: MemTensor Open-Sources MemT-1.0, a Memory-Native Foundation Model(11 posts)→
More from Research
- ICA Lens in action: decomposing Qwen residuals reveals clean city/France features — furongh · 2026-10-21
- Into the Unknown: none of 135 submissions by 9 frontier agents passes scientific bar — wellecks · 2026-10-21
- REUF2F Benchmarks Coding Agents on Formalized Undergraduate Research Problems — wellecks · 2026-10-21
- Baker Lab and NVIDIA Unveil RFD4-Proteina, a Faster Protein Design Model That Beats Prior Methods — jueseph · 2026-10-21
- NVIDIA Kumo Forecast Pairs Pretrained Time-Series Representations with Pattern Retrieval — Hugging Face Blog · 2026-10-21
- Protein designers adopt RFD4-Proteina for Adaptyv x Anthropic competition — AllThingsApx · 2026-10-21