What is actually inside an AI second brain?

People ask what a second brain is and get an answer about philosophy. Here is the answer about parts.

An AI second brain has four layers, and only one of them is difficult. It is not the one most people expect, and it is not the one any tutorial spends its time on. This post walks through all four, what every build comes with as standard, and what the whole thing actually runs on.

What is an AI second brain made of?

An AI second brain has four layers: the vault, the rules file, the agent, and the connections. The vault is a folder of plain text files on your own machine. The rules file tells the assistant how you work. The agent is an AI that lives inside the folder and can write to it. The connections are the tools it is allowed to read, added last.

Everything else is detail. If you understand those four, you understand the product.

The vault

A folder of plain text files, sitting on your own computer. Notes, decisions, people, processes, meeting records. Obsidian is the usual window onto it, but Obsidian is a window and not the thing. If it vanished tomorrow the folder would be fine, and so would you.

The rules file

One plain-English document at the root of the folder that the assistant reads in full before it does anything at all. Who you are. How your business works. Your customers, spelled correctly. What it may never do. This layer is the entire product, and it is the one nobody talks about.

The agent

An AI that lives inside the folder and can open, write and reorganise files itself, rather than telling you what to type and leaving you to do it. Most setups stop one step before this, and it is the step that matters. A system that can only read is a search box. A system that can write is an assistant.

The connections

Your calendar, your inbox, your task list, your documents. This is the layer everyone wants to start with, and it is the one to add last, one connection at a time, each one read-only first. Starting here is the single most common way a build goes wrong.

Why plain text files instead of an app?

Plain text is the format that outlives the tools. A markdown file opens in anything, weighs almost nothing, backs itself up the moment it sits in a cloud drive, and cannot be held hostage by a company that changes its pricing or shuts down.

There is a second reason that matters more as time passes. Plain text is the format an AI can actually read end to end. A proprietary database gives you an export button and a schema somebody else designed. A folder of text files gives you the thing itself, and any assistant you point at it can read every word.

That is the whole portability story. If you use one AI today and want to switch to another next year, you point the new one at the same folder and carry on. Nothing gets migrated, because nothing was ever locked up.

What is the rules file, and why is it the whole product?

The rules file is a plain-English document that the assistant reads before every conversation, and it is the difference between a clever demo and a system that is still running in three months. Folder structure is an afternoon of work. The rules file is months of catching an assistant being wrong and writing down why.

It holds the things you would tell a sharp new colleague in their first fortnight and then never repeat. How you price. Which customers are sensitive. The way you like an email to open. The fact that two people in your contacts have the same first name and confusing them would be embarrassing. What it must always ask about before doing.

The studio's own rules file is 889 lines long. It started at forty. Every line after the fortieth exists because something went wrong once and got written down so it could not go wrong twice.

This is also why a second brain cannot really be bought as a template. A template gives you the folders. It cannot give you the fortnight of corrections that turn a generic assistant into one that knows your business.

What does every build come with?

However different two businesses are, some pieces are standard. These are the parts that make the thing survive contact with a genuinely busy week, which is the only test that counts.

• The operating system. The rulebook it follows, written in plain English, in a file you can read and change yourself.
• The ingest lane. A place to drop anything: a document, a photo, a voice memo, a link. It gets read and filed without you filing it.
• The daily note. Where today lives. The plan on top, what actually happened underneath.
• The log. Everything it did, dated, so you can always see what changed and when.
• Your voice, on your phone. In practice you talk to it far more than you type to it.
• Your spaces, kept apart. Business and personal do not bleed into each other.
• The memory. A directory of the people you deal with, a journal of decisions and the reasoning behind them, and the list of things you have already said no to.
• Backups, twice. A folder of files backs itself up the moment it sits in a cloud drive, and again wherever you keep the second copy.
• The handover kit. A recovery guide, a list of what anyone else can reach, a map of where everything lives, and the setup itself written down.

That last one exists because a system only one person understands is a liability wearing the costume of an asset.

What does a second brain run on?

Four things, and you are probably paying for one of them already.

• Obsidian, or any markdown editor. The window onto your folder. Free.
• An AI assistant, on a paid plan. Claude, ChatGPT or Gemini all work. The habits are identical; the tool is whichever one you already use.
• A dictation app, so you can talk to it instead of typing. This turns out to matter more than people expect.
• A cloud drive, iCloud or Google Drive. This is also your backup, at no extra effort.

Running it costs roughly what a phone plan costs, and every dollar of it goes to companies that are not us. There is no software of ours to rent, no server, and no account with anybody new.

Do you need a vector database or a RAG pipeline?

No. For a business of one to twenty people, a vector database solves a problem you do not have. The assistant reads the actual files, the same way you would, and navigates a well-organised folder rather than guessing at fragments of one.

This is worth saying plainly because the category is full of architecture that exists to be impressive. Embeddings, retrieval pipelines and chunking strategies are real engineering and they matter at real scale. At your scale they add a layer that can break, that you cannot inspect, and that stands between you and your own words.

Five million words, well organised, is enough when the assistant can navigate rather than guess.

How long before any of this is useful?

It does real work on day one, because the first thing that happens is that it reads the documents you already have. It is genuinely compounding at around the three-month mark, once enough has been written down that it starts answering questions out of your own history rather than out of general knowledge.

The thing that changes at three months is not capability. It is that you stop explaining yourself. The questions you ask move from "how do I" to "what did we decide about", and only the second kind needs a second brain at all.

A note from Alex. The layer people want to talk about is the agent, because that is the part that feels like the future. The layer that decides whether you still use it in April is the rules file, because that is the part that makes it yours. If you only have attention for one of the four, spend it there.

If you want to see one running before you decide anything, the second brain build page explains what we do, and the longer field note shows a full working day on mine. Or just get in touch.