In this video I build an AI Second Brain the way I actually want to use it: locally, fast, and owned by me. I walk through the weekly challenge format we run in my community, then I take that same “build your own second brain” challenge and implement a simple version live with the YouTube crowd. The goal is straightforward: I want to chat to an agent in natural language, save different memory types (text, links, files/images), have the agent auto-tag and categorize everything, and then be able to search and recall it later—without losing data between sessions. I keep the stack practical: Next.js for the web app (frontend + backend), SQLite for persistence (because it’s perfect for local MVPs), and SQLite FTS5 for proper full-text search. For inference, I use the Claude Agent SDK (Opus 4.6, but you can swap to Sonnet) and I explain the key constraint: you don’t “assign tools” the usual way, so I expose database actions through an in-app MCP server instead. I also share how I think about context engineering—sketching the architecture first, then feeding docs and constraints to the agent so it can plan properly—plus a few opinions on why building your own harness matters if you don’t want to be locked into a vendor.