How to Build an AI Second Brain With Plain Markdown Files

What is an AI second brain?

An AI second brain is a memory and filing system for an AI assistant. Instead of re explaining yourself in every chat, you keep your context in files, and the assistant loads them first.

"Second brain" usually means a personal knowledge system. Here the reader is the AI, not just you. The same files also work for you, because they are readable notes.

Why does an AI assistant need a second brain?

Unless a tool has its own memory feature, an AI assistant starts each session fresh. It does not carry over yesterday's correction, your preferred format, or the decision you made last week.

So you either repeat yourself or you write it down once. Writing it down is the second brain. I use mine for accounting work.

What goes in an AI second brain?

I keep 2 core files at the top of the folder, and the split between them matters.

When something doesn't fit either test, the assistant suggests which file it belongs in and I confirm. Mixing rules and facts in 1 file is the fastest way to a messy brain.

Under those 2 files, each area of work gets its own folder with its own rulebook and memory. The top level file tells the assistant which folder to read for which job.

Why plain Markdown?

Markdown is plain text with light formatting. Any AI assistant that can open files can read it, and so can you, in any text editor.

That matters for 3 reasons:

• No lock in. The files work with any assistant that can open them, so you can switch tools and keep your brain.
• Easy to inspect. You can open a file and see exactly what the assistant knows.
• Easy to sync. Plain files sit in a normal folder, so a cloud drive can sync them.

I keep the folder in a synced drive and open it in Obsidian to browse and link notes. Obsidian only reads the same files, so there is no second copy to keep in sync.

How do you make the AI read the files every session?

You have 2 options, and they differ a lot.

Tools that can see your folder. Some coding assistants read a small stub file such as AGENTS.md automatically at the start of every session. The stub points the assistant to the handbook and the memory file, so there is nothing for you to do each session.

Chat apps that cannot see your folder. Custom instructions in chat apps such as ChatGPT are text only and cannot read your files. Check your own app, because features change. For apps like that, write a short self contained summary of the brain and paste it once into the global instructions. The full files then act as the depth layer when you upload them or use a project.

Plan for both. A summary is not the brain, so keep the files as the source of truth.

How to build an AI second brain with Markdown files

1. Create a folder and 2 files. Name them HANDBOOK.md and MEMORY.md. Put the folder where your cloud drive syncs.
2. Write the identity section first. Who you are, what you do, and how you want answers: short, direct, 1 recommendation. Keep it to what changes the assistant's behavior.
3. Add voice rules for anything written in your name. Tone, sentences, words you never use, words you do. Point the assistant to read this before every draft.
4. Make 1 folder per area of work. Each gets its own HANDBOOK.md and MEMORY.md. Add a routing table to the top file: "if the task is X, read folder Y."
5. Add the auto load stub. A small file such as AGENTS.md that points the assistant to HANDBOOK.md and MEMORY.md. For chat apps, paste a short summary into the global instructions.
6. Turn every correction into a rule. Details in the next section.
7. Write facts to the memory file when you say "remember this." The assistant saves it and confirms.
8. Prune and merge. Remove rules that no longer apply and merge duplicates. I use a maintenance agent that proposes those changes, and I approve each one before it goes in.

How do you write rules the AI actually follows?

Fixing a mistake in the moment feels done, but tomorrow's session, without a memory of its own, starts fresh and repeats it. So I stopped correcting and started writing.

Each rule has 3 parts:

1. The rule. What to do, in 1 line.
2. Why. The exact mistake that caused it, with the date.
3. How to apply. When it kicks in and what to check.

The "why" line matters most. It lets the assistant handle a new case the rule never named. If a rule needs a paragraph, it is really 2 rules.

1 rule I wrote: read every file in full, and if anything comes back cut off, stop and read it again before acting. A read that was cut off partway once made the assistant skip several of my rules. None of the rules were wrong. They were never loaded.

In short: when an AI breaks a rule, check whether it ever saw the rule.

What mistakes break an AI second brain?

• Rules and facts in 1 file. Behavior and status change at different speeds. Split them.
• A rulebook that grows without pruning. Old rules conflict with new ones, so prune and merge them.
• Relying on the summary alone. The pasted summary is a pointer. The files hold the detail.
• No human check. The AI follows the rules you wrote, not the ones you forgot. Keep a person in the loop for anything that matters.
• Storing secrets in the synced folder. My rule is that secrets and standalone code projects stay out of it.

Is there a starter kit?

I packaged the structure I use as a starter kit. The free version is Second Brain OS Core, an email signup on Gumroad. Your AI assistant interviews you and fills in the placeholders itself.

FAQ

What is an AI second brain?

A folder of plain Markdown files that an AI assistant reads at the start of each session. It holds your rules, your decisions, and your context, so you stop repeating yourself.

Do I need Obsidian to build an AI second brain?

No. The files are plain Markdown and work in any text editor. Obsidian is a convenient way to browse and link them, but it only reads the same files.

Will ChatGPT or another chat app remember my files?

Not on its own. Custom instructions in chat apps such as ChatGPT are text only and cannot read files, so paste a short summary there and keep the full files for tools or projects that can open them. Check your app, because features change.

What is the difference between the rulebook and the memory file?

The rulebook says how the assistant must behave: always, never, and before X do Y. The memory file stores facts that can change, such as status, decisions, and dates.

Where do I start?

With 2 files and a short identity section. The brain gets better as you turn each correction into a written rule.

What should I keep out of an AI second brain?

I keep secrets and standalone code projects out of the synced folder, and a person signs off before anything important goes out. Decide what your own folder may hold before you start.