Brain in a Folder

01What this is

Your agent reads one page of instructions, asks you a few questions, and builds a folder of notes about you and your work. After that, it opens the folder whenever you work together. It knows what you do, how you write, and the rules you care about, so you stop explaining yourself every time.

It is built on Jake Van Clief's ICM method (Interpretable Context Methodology): small folders with short notes that tell the AI where to look and what to do.

02What you get

A folder like this one. Yours will fit what you told it.

my-brain/
  CLAUDE.md
  AGENTS.md
  about-me.md
  where-my-stuff-lives.md
  log.md
  proposed-changes.md
  system/
    voice.md
    rules.md
  jobs/
    newsletter/
      CONTEXT.md
  raw/
  wiki/
  outputs/

• CLAUDE.md and AGENTS.md: where the AI starts reading.
• about-me.md: who you are and what you do.
• where-my-stuff-lives.md: your other files, apps and accounts.
• system/voice.md and system/rules.md: how you write, with real examples, and the small things you always want.
• jobs/: one folder per repeating job, each with a short CONTEXT.md that says what comes in, what goes out and who checks it.
• raw/, wiki/, outputs/: things you saved, notes the AI keeps up from them, and finished work.
• log.md and proposed-changes.md: every change, and edits waiting for your OK.

Every file is plain text. You can open and edit any of it yourself.

03How it works

1. A quick check. The agent makes sure it can save files. If it can't, it tells you how to fix that.
2. What is this brain for? You say it in a sentence or two.
3. Show it your writing. Paste a few things you wrote and sent, so it can learn your voice.
4. Six to eight questions. One at a time, about 10 minutes. Skip any you like.
5. See the plan. It shows you the folder before it builds anything. Nothing happens until you say "yes, build it".
6. It builds the folder.
7. One real task. You do something you actually need this week, together, using the brain.
8. A regular review. It sets up a short check, weekly or more often if you like, to file new things and suggest changes.

04Two kinds of brain

Many people end up with a mix. The agent picks the shape from what you tell it.

05You stay in charge

• The AI can add new notes on its own.
• It asks before changing anything you wrote, including your voice and rules.
• Every change gets a line in log.md.
• Nothing gets deleted.

06Is this for you?

It is a good fit if you use an AI agent most weeks and want it to know you without a long setup. A few rules that help, whatever you choose:

• Name the one job first. Most regret comes from building a library when you wanted an assistant.
• Start with what you can finish today. Add more only when something hurts.
• Keep files as the source of truth. Treat any service as an add-on. Services can close.
• Plain folders work everywhere. A folder with AGENTS.md and CLAUDE.md works in every agent.
• Don't design for a team until a second person actually uses it.
• Level up one step per real problem. Never because a feature list looks good.

When to use something else

Most tools add to this folder rather than replace it. They are listed under Level up later. Pick something else from the start only if:

• You want an app with buttons, not a folder of files. Khoj is one: an app that searches and chats over your documents.
• You only use ChatGPT on the web or a phone app. Those can't see a folder, so start with a shared memory database like Open Brain (see Level up later).
• You live in Obsidian and want a big kit of plugins and commands. This stays small on purpose.

07Which apps it works in

• Claude Code
• Codex
• Cursor
• The Claude desktop app, with a folder attached

Chat-only apps on the web or your phone can't save files. If you try one, the agent will tell you, and it can give you the files to paste in by hand.

08Your files stay yours

The brain is a plain folder on your computer. Keep it in your own iCloud, Google Drive or Dropbox folder if you want it on more than one device. There is no account and no server of ours, and nothing is sent to us.

Your AI app reads the files while you work, the same way it reads any file you give it. If you stop using AI, the notes are still there and still readable.

09Level up later

Start with folders. Add one tool per real problem, never per feature list. In week 1, add no extra tools.

Each tool below fixes one problem. Add it only after you have seen that problem yourself. Links checked on 2026-10-10.

~/level-up

• $claude-mem

  #when you keep re-explaining what happened last session

  #records each session and brings it back next time

  Needs: Node 20+, Bun and uv. For Claude Code and Codex

• $gbrain

  #when your notes reach the thousands and the agent misses ones you know exist, or two agents must share one memory

  #keeps facts with their sources and can import your folder as it is

  Needs: Bun, and a local database to start (no key)

• $open brain (ob1)

  #when you need the same memory in apps that can't see your folder, like ChatGPT on the web or your phone, or you want to capture notes from Slack

  #puts your memory in one database every AI app can reach

  Needs: a Supabase account, an API key, and about 45 minutes

• $karakeep

  #when saved links and screenshots are scattered across tabs and phones

  #one inbox for links, PDFs and images, which can feed raw/

  Needs: Docker and a computer that is always on

• $obsidian-skills

  #when you live in Obsidian and the agent breaks your wikilinks or frontmatter

  #teaches the agent how Obsidian notes work

  Needs: an agent that supports skills

• $basic-memory

  #when you want a chat app, like the Claude desktop app, to read and write your notes through a tool

  #connects a chat app to a folder of notes

  Needs: Python with uv, and an app that supports MCP

• $grill-with-ui

  #when you want a nicer, clickable interview next time

  #puts the questions on a page in your browser

  Needs: Claude Code and Node 20+

• $scheduled tasks

  #when the brain review works and you want it to run on its own, weekly, daily or on any schedule

  #your agent's own scheduler runs it for you, in the Claude desktop app or the Codex app

  Needs: the app open, and your computer awake, at the scheduled time

• $

10Credits

• The method is Jake Van Clief's ICM. Learn it from the source in his free Clief Notes lessons and his short How to Build an AI Second Brain for Your Business.
• The research brain (sources in raw/, AI-kept pages in wiki/, an index and a log) follows Andrej Karpathy's LLM Wiki idea.
• The interview style comes from Matt Pocock's grilling skill.
• The clickable interview idea comes from Jason Ku's grill-with-ui.

Brain in a Folder is an unofficial project. It is not made or endorsed by Jake Van Clief, Andrej Karpathy, Matt Pocock or Jason Ku.