An AI second brain with Claude Code and Obsidian: how I built and use it

I built a second brain, and I use it all day. My memory no longer sits only in my head, but in a system that remembers and works alongside me, for my business and beyond.

In this piece I show how I built it, what it does for me every day, and how you can set one up yourself. No theory from a whitepaper, just my own setup.

That system has two parts. The first is Obsidian, an app where you keep your notes as plain text. My whole collection of notes (Obsidian calls it a vault) is the memory: every client note, every meeting record, every idea and every decision lives in it, searchable. The second is Claude Code, Anthropic's AI assistant that runs on my own computer. Those are the hands: it reads that vault, writes to it and does work with it. Not a chat window that has forgotten by tomorrow what we discussed today, but an agent with permanent memory.

For me this is not a luxury. I have ADHD: without a system that holds on to what I let go of, half of it hits the floor. This second brain catches it.

The gap nobody names: memory

Most people use AI like a slot machine. One prompt, one answer, and tomorrow you start from zero again. The real problem with tools like ChatGPT is not the intelligence, but the lack of memory: the context disappears the moment you close the tab.

You do not need a complex technical setup for this (developers call it RAG). A vault the agent can read and write to is enough. Obsidian is plain text, local, mine. Not locked inside one vendor's memory. If a better model comes along next year, my brain simply moves with it.

But don't ChatGPT and Claude remember things?

True, and it is a fine first step. ChatGPT remembers things across chats, and Claude now does the same. For one-off tasks that genuinely helps. But the moment you use it seriously, you hit three walls:

• You do not know exactly what it remembers. The memory is a black box the vendor controls.
• It is tied to one model. Move to a better model or another tool, and your memory stays behind.
• It is not yours and not structured. You cannot search it, version it, or let your own agents work in it.

An owned vault flips that around: you see every line, it travels with you to any model, and your agents read and write to it. Built-in memory is where you start. A second brain is where you end up.

How I built it

It sounds big, but it is a handful of parts working together. Here they are, in the order they connect.

The vault is the foundation. One folder of plain-text files in Obsidian. Notes on clients, records of conversations, ideas, decisions, daily logs. No database, no lock-in. Just folders and markdown I can still open in ten years.

Claude Code is the engine. The CLI runs in that folder and may read and write. It knows how I work because I wrote it down in an instructions file (my CLAUDE.md): who I am, how I write, what is and isn't allowed. That file is the constitution. Every session starts there.

A memory layer on top. Separate memory files the agent updates, so it remembers across sessions what we decided earlier. Not a vendor's black box, but files I can read and correct myself.

A search layer over it. A local search engine (qmd) across the whole vault, by keyword and by meaning. That lets me ask my whole business a question instead of digging through folders.

Hands to the outside. Connectors via MCP, an open standard for connecting AI to other software, to the tools I use daily: my calendar, my mail and MacWhisper for transcription. That lets the agent not only think but act.

A second agent as reviewer. Before anything is final, a separate, critical agent checks the first one's work. That catches mistakes and inconsistencies before they go out.

Everything stays local

This is the part that matters most for a smaller business. My vault runs 100% locally. It sits as plain text on my own machine, not in a vendor's cloud. Private data and client information do not leave my computer unless I explicitly allow it. When Claude Code uses a file, that specific content goes to Anthropic's cloud model, and I approve that step each time. Everything else stays local.

And it can be stricter still. If I want nothing to leave at all, I run a local AI model on my own machine and everything stays in house. In practice I work hybrid: light tasks go to local models, and the heavy work goes to Claude Code's more powerful Opus model. That way I decide what stays local and what may go to a stronger model.

The recipe

You do not have to be a developer to do something with this. Here is the shortest route, from simple to powerful.

What you need:

• Obsidian, or really any folder of markdown files
• Claude Code, the CLI
• A CLAUDE.md with your context and your rules
• A transcription tool such as MacWhisper that writes conversations away automatically
• Optionally a local search layer such as qmd for "ask your whole business"
• A few fixed commands for recurring chores

The steps:

1. Make your vault. One folder, a few subfolders (clients, conversations, ideas, decisions, daily). Start small.
2. Point Claude Code at that folder and write your CLAUDE.md. Who you are, how you work, what the agent may and may not do. This is the most important part.
3. Add a memory layer so the agent remembers what carries across sessions.
4. Connect your transcription so conversations land in your vault without retyping.
5. Lay a search layer over the vault for the "ask your whole business" trick.
6. Build your first command. Start with an end-of-day save that files everything neatly.
7. Connect the tools you work in, like calendar and mail.
8. Keep improving. Every chore you do more than twice, you automate with a command.

What it does in practice

Four things I do with it every day.

A conversation becomes knowledge by itself. I record a client conversation. MacWhisper turns it into text automatically, and an agent pulls out the summary, decisions and action points and files it all in my vault. I retype nothing. A week later I still know exactly what we agreed, because it sits searchable in my brain.

My week of input becomes content. One command reads the videos and articles I saved that week and distils the useful insights. A second turns those insights into concrete content ideas. What I consume becomes raw material for what I make.

I run my business through it. My calendar, my mail and my social posts run through that same brain. The agent proposes, I approve or adjust. I almost never type an answer from scratch. That is the difference between a chatbot and an AI agent that takes work off your hands.

I ask my whole business anything. "What did we agree with that client in March?" "What is the status of that proposal?" Instead of searching folders and mail, I ask the question and the answer comes from everything I have ever recorded. My business has become searchable.

DataDreamOS: my business runs on it

The vault started as memory, but it grew into the operating system of my business. I call it DataDreamOS. My whole administration lives in it, and the agents work with it directly.

A few examples from practice:

• I create invoices from the data already in the vault, not in a separate accounting package I feed by hand.
• I track my finances, leads and sales pipeline in it: who is in which stage, what is outstanding, what is coming in.
• For each client and contact I have one overview: every conversation, every appointment and every open action in one place.
• Action points and tasks roll out of my conversations and notes automatically, so nothing slips through.

It is not a set of separate tools side by side, but one system where everything comes together. That is the difference between loose AI tricks and a business that genuinely runs on AI.

This is agentic engineering, not vibe-coding

Agentic engineering means using AI like an engineer: with structure and control, not by luck. Here is the difference that matters. Refuse AI and you fall behind. But let AI loose without discipline and you ship fragile work that collapses at the first edge case. The optimum sits in between: an engineer's discipline, carried out through AI.

In practice that means structure first, execution second. I set out in advance what is and isn't allowed, what the output should look like, and which steps come in which order. The prompt is not the work, the prompt is the result of a thought-through workflow. Only once that scaffold stands do I let the agent run.

Vibe-coding is typing a prompt and hoping. Agentic engineering is building a system you can optimise, repeat and hand off.

What this means for you

A founder I work for summed it up recently: I keep him ahead with AI without it costing him time. That is exactly what a second brain delivers.

You do not have to rebuild my setup. The principle is what counts: make sure your context is yours and sits in one searchable place, and map out your workflow before you let AI loose on it. A business that has its knowledge scattered across six systems and throws AI loosely over the top gets loose tricks. A business that has its memory in order and its workflow mapped out gets a second brain that works alongside it.

And this is not just for businesses. The same principle works just as well for your personal life: your plans, your notes, your admin, everything that now disappears into scattered apps and your head.

That is the difference between AI as a toy and AI as a colleague that never forgets anything. If you want to know what that looks like for your processes, that is exactly the work I do: AI solutions that run in production.