Qwen 3.5 MoE + TurboQuant + mem0: A Local RAG Chatbot That Remembers

I upgraded my local RAG chatbot to a 35B Qwen mixture-of-experts at 32K context, running entirely on a MacBook. The trick is TheTom's TurboQuant fork of llama.cpp, which compresses the KV cache by ~60% with no measurable quality loss. This time it does real work: RAG over 94 full IEA energy reports, plus persistent cross-session memory with mem0: the bot remembers who I am after a full restart. mem0 is an open-source memory layer for AI apps. It works with your existing stack, local or hosted: run it on your own machine while you build, plug into managed memory when you ship. 👉 Get started: https://mem0.ai/?via=julien 👉 Promo code JULIEN: $19 off all payments within the first 3 months! The chatbot + mem0 integration & launch scripts: https://github.com/juliensimon/local-rag-chatbot/tree/feature/mem0 ⭐️⭐️⭐️ More content on Substack at https://www.airealist.ai ⭐️⭐️⭐️