The short answer: A second brain with AI is one external place to keep your notes plus an AI layer that reads, connects, and surfaces them, so you stop being the retrieval system. Build it in this order: pick one home for your notes, write in one daily file, then add an AI step that reviews that file. Everything after that is scale. Mine holds more than 4,000 notes and the AI tends all of them.
The part nobody tells you about note systems
I kept notes for years before any of it worked.
The notes were fine. The problem was that I was still the index. Every time I needed something, I had to remember that I had written it, remember roughly where it lived, and go dig. The system stored well and retrieved badly. Which meant it only worked when my memory was already working.
That is the wrong dependency. A system you have to remember to use is not doing the work you built it to do.
What changed was not a better folder structure. It was putting something in front of the notes that could read all of them at once.
Older systems stored. AI retrieves, connects, and explains. That is the whole difference, and it is why this is worth building now rather than five years ago.
What a second brain with AI actually is
A second brain is an external system that holds the information, context, and connections that would otherwise live only in your head. Tiago Forte popularized the term. The practice is much older than the term.
A second brain with AI adds one capability the earlier versions never had. The system can tend itself. AI reads what you wrote today, links it to something you wrote six months ago, summarizes it, and hands it back when it becomes relevant. You do not do the filing.
You already run early versions of this. A calendar. A notes app. A task list. Those are all external memory. What changes is that the system stops being a drawer you dig through and starts behaving like something that hands you the right note at the right time.
If you want to see what that looks like fully assembled before you build your own, I walked through the complete architecture in My Brain Has a Backup.
How to build a second brain with AI
You do not need 4,000 notes to get value out of this. You need one file and a habit. The order matters more than the tools.
1. Pick one home for everything
One place you will actually return to. I use Obsidian, which stores notes as plain text files on my own machine. Notion, Apple Notes, or a single document all work. Picking one and staying there beats picking the best one.
2. Start with one daily file
Call it whatever you will reopen. At the end of the day, write three things: what happened, what you noticed, what is still open. That is the entire habit on day one.
Three lines a day sounds too small to matter. It is not. The daily file is the only input the rest of the system needs, and it is the piece most people skip straight past on their way to building something more impressive. More on how those entries stack up over time in how my note system compounds.
3. Add the AI review step
Paste the day’s entry into an AI tool and ask it something like this:
Here is today’s log. Tell me the most important thing to carry into tomorrow, the one open loop I need to close, and one thing I learned today even if I did not call it a lesson.
That third clause does most of the work. People are reliably bad at noticing their own lessons in real time, and an outside reader is not.
4. Give it context, not just questions
Tell the AI who you are, what you are building, and what matters this season. Persistent context is what separates a chatbot from a second brain. Without it you get generic answers to specific problems, which is the most common reason people try this once and conclude it does not work.
Andrej Karpathy’s LLM wiki approach is the cleanest version of this I have found, and I adapted it directly into my own vault. I wrote up how I use that setup in Obsidian.
5. Build the two anchors
An end-of-day reset and a morning brief. Those two routines are the backbone. Everything else is optional.
The reset pulls from my calendar, email, and notes, writes a summary, flags the open loops, and appends to a running log. The brief is assembled before I am awake and holds my priorities for the day and what I said mattered this year. I broke down the morning system separately, and the assistant that runs both in more detail.
Start with one of the two. The morning brief is the easier win because you feel it the same day.
6. Review weekly and let it compound
Once a week, read back through the entries. The value was never any single note. It is the pattern you can finally see because something else held the pieces while you were busy living through them.
The stack I actually run
Two pieces.
- Obsidian holds the vault. Local, plain text, free, mine.
- Claude is the layer that reads, connects, and surfaces.
Swap either one. ChatGPT or Gemini work as the intelligence. Notion or Apple Notes work as the store. The architecture is what matters, not the brand. One place to write. One intelligence to tend it.
Cost is the one thing worth watching once the vault gets large, because a system that reads everything every time gets expensive fast. I cut mine by about 60 percent without losing capability, and wrote up how that worked.
What the AI layer is good at, and what it is not
Worth being honest about, because the gap is where people get burned.
It is very good at connection. Ask it what today’s entry has in common with anything from the last two years and it will find threads you would not have gone looking for. It is good at summarizing, at tagging, and at answering questions about your own material.
It is not reliable as a source of truth about your life. It will state something with total confidence that your notes do not support. I have watched it happen more than once, and the fix is structural rather than clever: keep the notes as the record and treat the AI as the reader, never the other way around. The store is the memory. The model is just the interface.
Where most builds fail
Building the system before the habit. A perfect vault with no entries in it is a folder. Start ugly.
Treating it as storage. If you file things and never ask the AI to surface them, you have rebuilt a filing cabinet with extra steps. Retrieval is the point.
Starting cold every session. Give the AI your context once and reuse it. Reconstruction is the expensive part, and removing it is the entire reason to build this.
The short version
- One external store plus one AI layer that connects and surfaces. That is the whole architecture.
- Start with one daily file and one review prompt. Do not build the system first.
- Two anchors carry it: an end-of-day reset and a morning brief.
- Persistent context is the difference between a chatbot and a second brain.
- The tools are interchangeable. The architecture is not.
Blake Murphy builds AI systems and writes about leverage, knowledge systems, and thinking clearly. The companion piece is My Brain Has a Backup, which walks through the full architecture of the system described here.