MemT-1.0 Nano: MIT-licensed 36B/3B LLM embeds persistent memory inside the model
solyarisoftware · x · 2026-10-21
New open model MemT-1.0 Nano takes a different memory approach: instead of stuffing old conversations back into context, memory lives inside the model — memory blocks run alongside attention, learning to write, forget, update and read information. Built from Qwen3.6-35B-A3B (36B total / 3B active), with 10M dynamic memory state and 451M trainable memory parameters, MIT-licensed. Authors claim persistent memory is 66× smaller than a full KV cache at 128K history. Benchmark charts show its parametric memory beating MemOS, DenseRAG, Graphiti, Doc-to-LoRA, SHINE, 6-mem and MemoryLLM across LoCoMo, MemoryAgentBench, ImplicitMemBench, MemBench, PerMemSafe and PersonaMem.
Related event: MemTensor Open-Sources MemT-1.0, a Memory-Native Foundation Model(9 posts)→
More from Models
- Uncensored DeepSeek V4.1 Flash GGUFs drop refusal rate from 91.4% to 0% with minimal MMLU loss — kalyan_kpl · 2026-10-21
- Meta engineer confirms Google TPUs now run AI inference in Meta production — dotey · 2026-10-21
- Don't mention recent AI math results to Opus 5.5 if you want it to focus — repligate · 2026-10-21
- User finds codex now beats ChatGPT for everything but quick phone searches — verrsane · 2026-10-21
- Nace.AI Launches Drex Max 2.0, a 27B Local Decision Model for Computer Use — DotaMate · 2026-10-21
- Speculation: Codex/Work also cap at 12 hours — OpenAI quietly eats cache-miss costs — dotey · 2026-10-21