Personal AI brain
Running every day. Cloud routines run weekly and every weekday morning.
One folder of plain text files that every AI I use reads before it does anything, so Claude Code, Codex and Antigravity all know me and my rules without me explaining twice. It also runs its own weekly routines to keep itself tidy and up to date.
- Why
- Every AI tool kept its own private memory of me, and none of it came with me when I switched tools. I wanted one memory that I own and any AI can read and write. Now a new tool starts with everything the old one knew: in July I pointed Codex and Antigravity at the folder and they gave me answers as good as Claude Code on day one.
- How long
- First version over a weekend in March 2026. It has grown almost every week since: 700+ commits in seven months.
How I built it
- I made a vault in Obsidian (free at obsidian.md), which is just a folder of Markdown files. At the top I put a README that tells any AI: read the _AI folder first, then check the task queue. Tiny CLAUDE.md and AGENTS.md files in the same folder only say "read the README", so every tool lands in the same place.
- Inside _AI I wrote the rules: three core files (about me, my writing voice, how I like to work), a memory policy that says this folder is the only memory, and write rules. The AI can edit regular pages freely and tell me, but has to show me any change to a core file first.
- I put the folder in a private GitHub repo. In Claude Code I asked it to set up git, create the repo with the GitHub CLI and push. Now the AI commits and pushes after every change, so I get sync, history and an undo button.
- I added a Task Queue page for open threads, so I can stop on my laptop and pick up on my phone, or in a different AI tool, without re-explaining.
- I set up Claude Code cloud routines (claude.ai/code, Routines). Each one is a short pointer to an instruction page in the folder: a weekly recap journal on Monday evenings, a weekly cleanup that fixes broken links, a weekday email triage and a 5:30am briefing. To change a routine I edit the page, never the routine.
- When the AI gets something wrong, it writes the fix into the folder as a rule or a step-by-step SOP, so the next session does not repeat it. That is how it grew to about 15 SOPs.
- To copy it you need
- Claude, or the AI of your choice (Claude Code, Codex and Antigravity all work, because it is just files). Obsidian (free), a GitHub account with a private repo, and an hour. Gotcha: Claude and ChatGPT will happily save facts about you in their own built-in memory. Tell your AI plainly to write to the folder instead, or you end up with two memories that disagree.
Try it yourself
Copy this prompt and paste it into ChatGPT or Claude to try it yourself.
I want to build a personal AI brain like Amelia Lin's (it's on Amelia's What Are We Building? desktop at whatarewebuilding.ai). It's an Obsidian folder of plain Markdown files that every AI tool reads first. A README says: read the _AI folder before anything else, then check a Task Queue page. _AI holds three core files (about me, my writing voice, how I like to work) and a memory policy: this folder is the only memory, so never save facts about me in your own built-in memory. Start small: help me create the folder, the README, and a first draft of the about-me file from what you already know about me, marking anything you are guessing. Then help me put it in a private GitHub repo. The gotcha that bit Amelia: AI sessions that save work to side branches nobody merges, so pages quietly vanish. Always commit to main. Ask me for anything you need.
Tools I used for this project










