I spent a month testing an AI second brain
# I spent a month testing an AI second brain
The second brain pitch is everywhere at the moment. Dump everything you write into one place, all the notes and links and half thoughts from two in the morning, and an AI recalls it on demand, connects ideas you forgot you had, and quietly thinks alongside you. I wanted to know what that promise looks like after the demo ends, so I ran it properly for a month. Some of it held up. Some of it fell apart in ways I found more interesting than the failures I expected. And one lesson landed on me personally rather than on the tooling.
Why the pitch works on people like me
The problem this category claims to solve is real, and I have it. Most people take notes and never look at them again. You write something down, feel the satisfaction of having captured it, and it disappears into a folder you open once a year. The second brain argument says the failure is in retrieval. If you could question your notes in plain language and get a real answer back, you would actually use them.
The pitch goes further than search. The AI is supposed to spot connections between things written months apart, surface a three week old note exactly when it becomes relevant, and behave like a memory layer that needs no tending. I had years of raw material to test against: voice memo transcripts, half finished drafts, saved articles, phone notes typed at odd hours. If the idea worked anywhere, it should work on that pile.
What I actually set up
I want to be specific, because vague promises are easy to chase and hard to test. I spent about a week moving everything into a single folder an AI model could read. Text files, cleaned up voice memo transcripts, short notes, longer drafts, saved ideas. I deliberately left the notes messy, since handling raw imperfect input is the entire point of the category. Tidying them first would have been grading the tool on a curve.
Then I lived with it for four weeks. I kept writing new notes the way I always had, and whenever I wanted to recall something or think something through, I went to the AI before searching manually. The model behind it was one I pay for, which is worth naming, because nothing serious in this category runs at zero cost. The cost turned out to be the least interesting part of the results.
Plain language search earned its keep
The clearest win was the simplest one. I could ask what I thought about a topic three weeks ago, or whether I had any notes on something, and the relevant passage came back. That sounds minor until you know my habits. I have a long standing pattern of writing something, losing it, and rewriting a slightly worse version of the same thought a month later. The search caught several of those moments in the act. It surfaced notes I had completely forgotten writing.
It missed things too. Recall was never total. But it found more than I would have found by scrolling, and the lower friction meant I consulted my own notes far more often than before. If this had been the only result, the experiment would still have come out positive.
Summarising the mess held up too
The second win was distillation. My notes tend to come out as long streams of consciousness that never get boiled down. Pasting one of those rambles in and asking for the main points gave me something readable in seconds, and it nearly always caught the core idea even when the surrounding text was incoherent.
This one surprised me. I had assumed summarising my own thinking would feel redundant, since the thoughts were already mine. Instead it felt like having a patient reader who worked through the mess so I could skip to what mattered. By the second week I was using it routinely, and it is one of the pieces I still use now that the experiment is over.
The connections were rare and real
The third result was the one the marketing leans on hardest. Occasionally, while I was asking about one topic, the AI pulled up a note from a completely different context that turned out to matter. Over the whole month this happened maybe six or seven times. Nothing like daily. But those moments were the most interesting part of the entire exercise, because they demonstrated a kind of usefulness a search engine cannot offer. A search engine needs you to know what to look for. These surfacings required no query I could have thought to write. A few of them genuinely changed how completely I thought something through.
So the headline promise is partially real. Partially is doing a lot of work in that sentence, which brings me to the failures.
It invented things with a straight face
The most uncomfortable discovery was that the AI sometimes manufactured connections. I would ask whether I had written about an idea, and it would say yes and summarise notes that, when I actually opened them, said something different from the summary. In a few cases the described note barely existed at all. It had taken a real file and dressed it in claims the file never made.
I caught these because I checked. The frequency was low. The confidence was the frightening part. There was no hedge, no flag, no difference in tone between an accurate recall and an invented one. Had I trusted the output without verifying, I would have been building on foundations my own notes never laid, without any signal that it was happening.
For a tool whose whole job is extending your memory, this is a serious flaw. A memory that occasionally invents content, delivered with full confidence, can leave you worse off than no external memory at all, because at least an absent memory sends you back to the source.
The maintenance tax nobody mentions
The second failure was mundane. Feeding the system was work. For the recall to function, notes have to reach the folder in usable form, which means moving, naming, formatting, and keeping up with new material as it arrives. My note taking happens in inconvenient moments, on my phone while walking, in a voice memo while cooking, in messages I send myself. Converting that stream into clean input was a steady tax, paid weekly.
The pitch says the second brain frees you from manual work. In practice it relocated the work. I stopped manually searching my notes and started manually maintaining the system that let the AI search them. Totalled up, the hours were roughly the same as before. The work had changed shape without shrinking.
The lesson that was about me
The part that stuck longest had nothing to do with the software. A few times during the month I caught myself waiting for the AI to tell me what I thought. I would have a vague sense of having worked through an idea before, and instead of sitting down and reconstructing the reasoning, I asked the machine.
Sometimes that was fine. Sometimes it was quietly corrosive, because reconstructing the reasoning was the point. Doing the thinking again is what makes a conclusion stick and develop. Outsourcing recall is a reasonable trade. Outsourcing the thinking itself is a different transaction dressed in the same interface, and I had not noticed how easily the two blur. A second brain is sold as a memory tool, and it is remarkably easy to use it as a way to avoid the blank page. Notes produced that way get shallower, because nothing real happened before they were written.
The lighter version I kept
I did not throw the setup away. I kept the plain language search, because it works. I kept summarisation of long messy drafts, because it saves real time. What I dropped was the ambition of a live system holding everything, because that ambition generated most of the friction and nearly all of the disappointment.
The shape that survived is on demand rather than always on. When I have a batch of messy notes on one topic, I gather them and let the AI help me work through them. When a long draft needs distilling, I hand it over. Nothing gets fed continuously, and nothing depends on the system being complete. As a concept it is much less exciting. As a tool it gets used every week, which is the measure I care about.
Where I landed on the hype
My verdict is that the second brain is a genuine idea wearing too much marketing. Better retrieval over your own thinking is a real improvement, and the demand for it is honest, because almost everyone sits on notes they never use. The living, connected, accurate memory extension is ahead of what the tools reliably deliver right now. Invented connections are a real problem. The upkeep is a real problem. And the hardest problem belongs to the user, because a tool that answers instantly makes skipping the thinking feel free when it is anything but.
If you want to try it, start narrow. Pick one category of notes you actually revisit, use the AI to search and summarise within that category only, and verify what it returns before you lean on it. Let the habit come before the architecture, because a complicated system with no habit behind it is pure friction, while a simple habit with a little AI support gets used. That is the version that survived my month, and the version I would hand anyone starting fresh. For more honest accounts of what AI tools actually do in real use, I write them up regularly on [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.