A new open LLM can remember an old conversation without putting that conversation back into its context. And at 128K history, its authors claim the persistent memory is ~66× smaller than keeping the full KV cache. MemT-1.0 Nano dropped today. Stats 🧠 36B total / 3B active ⚙️ built from Qwen3.6-35B-A3B 💾 10M dynamic memory state 🧠 451M trainable memory parameters 🔓 MIT-licensed weights But it does things a little differently. Normally, a long-running LLM needs old conversation + new message. So as history grows, you keep dragging more context/KV around, or retrieving old text and stuffing it back into the prompt. MemT puts memory inside the model itself. Its memory blocks run alongside the normal LLM attention. As you interact with it, the model learns to: 📝 write information 🗑️ forget information 🔄 update information 🧠 read it back later So you can: 💬 have a conversation 💾 save its latent memory state 🗑️ discard the plaintext conversation 🔄 start a new session with zero previous messages 🧠 restore the memory ➡️ and ask about what happened before Their demo does this. At 128K history, the authors report for MemT-Nano 📦 persistent memory → ~66× smaller than full-context KV cache ⚡ end-to-end inference → ~1.86× faster And the memory state stays a fixed size as history grows. MemT isn't a little memory model sitting next to Qwen. It takes a normal generative LLM and makes persistent memory part of the LLM's computation itself. They also released MemT-Flash: 🧠 288B total / 17B active 💾 336M dynamic memory state 🧠 3.61B trainable memory parameters 🎯 This could be especially interesting for long-running local agents. ⚠️ The 66× / 1.86× results are author-reported, and strong text-retrieval systems still beat MemT on some tasks where exact historical details matter. I can't find consumer GGUF benchmarks yet. Link in ALT.