I got tired of AI assistants that forget everything the moment a session ends. So I built one that doesn't. It runs 24/7 on my Mac, has access to my files, GitHub, iMessage, email, and calendar. It knows who I am, what I'm working on, and what I said to it last week. Every time you open a new Claude or ChatGPT session, you start from zero. You re-explain your context. You re-establish what you're working on. You paste in the same background info. This is fine for one-off tasks. It's terrible for an ongoing working relationship. Instead of in-context memory, I use files: MEMORY.md — long-term curated knowledge. What matters, distilled. memory/YYYY-MM-DD.md — daily logs. What happened, decisions made, things to remember. USER.md — who I am, my stack, my communication style. TOOLS.md — local setup specifics. Every session, the agent reads the relevant files before doing anything. This is the continuity layer. Model Context Protocol (MCP) is what lets the agent actually do things — not just talk about them. I use it for: Apple Mail, Calendar, Messages via a local MCP server GitHub via gh CLI File system access Browser automation (Puppeteer via Chrome DevTools Protocol) It's not a chatbot. It's closer to a part-time assistant who's always available and never forgets anything. The most useful thing isn't any single capability — it's that context persists. I can say "remember the JWT issue from last week" and it actually knows what I mean. The hardest part isn't the AI. It's designing the memory and context system that makes it feel coherent over time.