Every Local AI I Run Now Shares ONE Memory | (LLM Wiki + OKF)
I gave every local AI I run ONE shared, permanent memory — teach something to one agent, and every other agent already knows it. No vector database, no cloud API, no re-embedding. Just a folder of markdown files, a "librarian" agent, and MCP. It's built on Andrej Karpathy's LLM-wiki idea and Google's OKF (Open Knowledge Format) spec, and the whole loop runs on my own box: the harness, the memory, AND the model (same llama.cpp server everything else uses). In this video I build understory, a drop-in memory layer for ANY local AI setup. AnythingLLM, llama.cpp web UI, coding agents, anything that speaks MCP, and I show you exactly what broke along the way: the cold-start problem (the AI never thought to check its own memory), the junk-drawer problem (every fact became an orphan file), and how a deterministic harness makes it all work on a small local model. Deterministic rules, the LLM only for decisions. 🌱 understory (free & open source): https://github.com/thecodacus/understory Karpathy's LLM Wiki gist: https://gist.github.com/karpathy/442a... Google's OKF spec: https://github.com/GoogleCloudPlatfor... ⏱️ CHAPTERS 0:00 Your local AI has amnesia 1:25 Why RAG is NOT the answer (memory ≠ vector DB) 3:04 The naive fix: a skill file (and how it broke) 3:50 Every library has a librarian — the architecture 5:26 Meet understory: watch the memory grow 5:59 Plug it into anything: the 3 MCP tools 6:53 Truly local, it runs on the same llama.cpp 7:44 Challenge 1: it forgot to remember (cold start) 9:19 Challenge 2: the junk drawer (linking memories) 11:01 Fixing the final gaps (lint, maintain, contradictions) 12:34 One memory, every agent 12:58 My Thoughts + the roadmap THE STACK • Memory: plain markdown files (Google's OKF spec) — no database, every "memory" is a file you can read, edit, and git-diff • Agent: TypeScript + Vercel AI SDK v5 tool loop (search / read / write / patch / link) • Protocol: MCP server (streamable HTTP + stdio) — works with Claude, llama.cpp webui, any MCP client • Inference: llama.cpp llama-server on my homelab GPU (also supports Anthropic / OpenRouter) • UI: React + Vite + Tailwind, d3-force for the memory graph + query-path replay • Server: Node + Express, single Docker container (image on GHCR) Every byte of your AI's memory stays on your disk, readable in any text editor, diffable in git. We took back the model, the hardware, and the privacy — memory is the next thing to take back. 🔔 Subscribe if you're building a local AI stack you actually own: / @codacus What should your AI remember first? Tell me in the comments. #localai #llamacpp #edgeai #aimemory #llm #obsidian #okf

Running a 35B AI Model on 6GB VRAM, FAST (llama.cpp Guide)

I Asked Claude Fable 5 to Improve llama.cpp.. and It Did

Finally, an Open Standard for the Karpathy LLM Wiki is HERE

The Best Local Agentic Coding Workflow (Complete Guide)

Build Powerful Local Coding Agent on Budget GPU with Llama.cpp and Pi

Cactus Needle - The 26M Function Calling Model

I Love the Karpathy LLM Wiki but it Doesn't Scale. Here's What Does.

Ornith 35B Benchmarked vs Qwen 35B - 16GB Local LLM setup

New GitHub Repos That Feel ILLEGAL To Get Free

SQLite + AI: The Database Revolution Nobody is Talking About (No Pinecone)

"Software Fundamentals Matter More Than Ever" — Matt Pocock

Everything That Actually Matters for Local AI

The Mitochondria Doctor: This Reverses Gray Hair, Makes You Feel Young Again & Fixes Disease!

Recursive Self-Improvement

The Hard Drive Cartel | Criminal Conspiracy & Price Fixing

China Is About To Pop The AI Bubble

How does AI actually work? Transformers explained

Claude 6: Anthropic's Next Move Shocked Everyone

Build Your Own Fully Private, Local AI Stack (Chat, RAG, Coding Agent, Automation)

