A second brain for Claude Code, and where it stops

What is a Claude Code second brain?

Three layers, all plain files. The notes are the markdown pages, written by you, by Claude, or both. The schema is a CLAUDE.md (or AGENTS.md) at the root that says where things go, how pages are named and what Claude should do with a new source. The viewer is Obsidian, which shows the same files as linked pages. Nothing needs a plugin or a server: Claude Code works on the folder with its normal file tools.

It is the same shape as Karpathy's LLM wiki, which is where much of the 2026 interest came from: the agent files what it learns once, cross-links it, and answers from the pages instead of rediscovering the same facts every session.

What does Claude Code load from the vault?

Less than people assume, and in a fixed order. From Anthropic's memory docs:

So the root CLAUDE.md is the one page you can count on being read. Put the map there: the folders, what each holds, and which index page to open first. Everything else is a page Claude has to choose to read. Anthropic suggests keeping each CLAUDE.md under 200 lines; imports organize a long file but do not make it cheaper, because imported files load with it.

How do I set one up?

• Make the folder. Open it as a vault in Obsidian if you want to read it there, and start claude in it.
• Write the root CLAUDE.md. The folder layout, a naming rule for pages, and what to do with a new source: summarize it, update the pages it touches, add a line to a log.
• Keep one index page. A list of every page with one line each. Claude opens it first and decides from it what else to read.
• Keep it shallow. A few top folders beat a deep tree, because Claude has to find a page before it can use it.
• Put the vault in git. It gives you history, and a way to undo an edit Claude should not have made.

Is auto memory already a second brain?

Partly. Claude Code writes its own notes to ~/.claude/projects/<project>/memory/ and loads the start of MEMORY.md into every session. It is a good memory of how you like to work. It is also machine-local: the docs say the files are not shared across machines or cloud environments, and only Claude Code reads them. A vault you choose to sync goes further, but it is still yours alone.

Where a personal second brain stops

Take one decision. In #eng-payments the team agrees that webhook handlers return a 200 and enqueue, and never do the work inline. You write it into your vault. Then:

• Your teammate's agent never sees it. The note is on your machine, in your vault.
• Other agents read other files. Codex reads AGENTS.md and never CLAUDE.md, and Cursor and Copilot each have rule files of their own. A Claude-shaped vault reaches Claude. The AGENTS.md vs CLAUDE.md note covers which tool loads what.
• Nothing gates what goes in. Alone, you are the gate. Shared, one wrong page is repeated by every agent that reads it; see agent memory poisoning.
• Nothing tells you a page went stale. When the queue is replaced, the old page stays until somebody remembers it.
• Nobody knows which pages are used. You can see what is in the vault, not what any session actually read.

A second brain for a team's agents

This is what Harbor is for. It is shared memory for engineering teams using coding agents: one store that Claude Code, Codex, Cursor and GitHub Copilot all read, wired with npx harborloop init. It learns from your sessions and, if you connect them, from Slack threads, pull request reviews and docs. You set the gate: routine rules can go live on trial, and a secret, an install step or a clash Harbor cannot settle waits for a person. Each session gets the rules that fit its task, not the whole store, and every rule shows how often it was served and how often an answer cited it. A cite is evidence a rule was used, not proof it was obeyed.

Keep the vault for your own notes, papers and drafts. It is very good at that. Put what the team decided where every agent on the team can get it.