My Second Brain Doesn’t Need to Know Me
I’ve been building what I loosely think of as a second brain. I’m deliberately not trying to make it remember everything about me.
That runs against most of the current advice. The assumption in almost every conversation about AI and personal knowledge is that more context produces more usefulness: feed the model your preferences, your history, your working style, your archives, and it will finally know you.
I think the more useful question is narrower. Not how much context to give an AI, but which context is worth keeping and where each kind of context should live.
The Library Holds Processed Thinking, Not Sources
The center of the system is a Learning Library in Notion. When I find something worth keeping (a webinar, article, podcast, research report, YouTube video, book, or one of my own frameworks) I don’t bookmark it. I run it through a gate.
I hand the source to ChatGPT, which evaluates it against a curation standard I defined up front. Most things don’t clear it. What does gets distilled into the fields I want to retain:
Core idea. What was this really about?
Key insights. What is worth remembering?
What changed my thinking. Did this challenge, sharpen, or reinforce something I already believed?
Connections. How does it relate to other things I’ve learned, frameworks I use, or problems I’m working on?
Applications. What could I do differently because of this?
Content ideas. Is there an argument here worth developing?
Questions to explore. What still doesn’t make sense?
The raw source can live somewhere else. What lands in the library is the compressed version.
That last field matters more than it looks. A note recording what I still don’t understand is the one I come back to when a later source finally answers it.
The template is the smaller half of this. The gate is what keeps the library worth reading. A knowledge system without an entry standard becomes a folder of things you meant to read, and the reason most of them collapse isn’t a bad tagging scheme. It’s that everything gets in.
Three Kinds of Memory Doing Three Different Jobs
This is where my thinking has changed.
I don’t need the Learning Library to remember that I prefer a certain writing style, what I’m working on this week, or the context of a conversation I had yesterday. AI memory and project context handle that kind of continuity reasonably well now, and they keep getting better at it.
The library has a different job. It holds the things I’ve deliberately decided are worth preserving.
That separation makes the whole system simpler:
AI memory knows me.
Project context knows what I’m working on.
The Learning Library knows what I’ve learned.
Most second-brain advice collapses all three into one undifferentiated pile, which is the same instinct that turns a CRM into a junk drawer: everything gets bolted onto the system people already log into, and nothing has a clear owner. Separating the jobs makes each layer easier to govern, easier to replace, and easier to trust.
The Library Has to Live Outside the Model
The library stays in a format I control, outside any single vendor.
That isn’t a prediction about which model wins. It’s a decision about where the durable asset sits. The storage is portable. The contents still carry one model’s judgment about what was worth keeping, which is a real limit on how model-independent this actually is. Today a model can reason across the library through a connector or an export. Whatever I’m using in three years should be able to do the same, and if it can’t, I move the library rather than rebuild it.
The model is replaceable. The accumulated knowledge isn’t.
Connections Are What Compound
The unexpected part is that the library has started behaving less like storage and more like a feedback loop.
I’ll process an article about AI-native marketing. Another source introduces a related idea about first-party data. A webinar shows how a Marketing Operations team is actually implementing it. Now there’s a connection across three sources that none of them made on its own.
That connection might become a framework, a workflow, a post, or just a better way to think about a problem six months from now.
None of that happens if I’m only collecting links. It happens because I wrote down what each source changed, and the writing forced the comparison.
Where This Breaks Down
I don’t want to oversell a system I’m still running.
The library is only as good as the standard I wrote. A vague bar produces inconsistent gatekeeping, and I won’t notice the drift from inside – I’ll just have a library that feels fine and quietly stopped being selective. Automation amplifies the quality of the underlying process, and that applies to my filter as much as to anyone’s routing logic. Periodically reviewing what got rejected is probably more useful than reviewing what got in.
The distillation is a compression, and I’m trusting a model’s judgment about what mattered in the source. Whatever it dropped, I don’t see again unless I go back to the original. That’s an acceptable trade for the first ten sources. I’m less sure it’s acceptable at two hundred.
Curation also cuts both ways. A standard built from what I already consider valuable is a record of my existing frame, now enforced automatically rather than reconsidered case by case. The “what changed my thinking” field is a partial defense, but only if I’m honest when the answer is nothing.
And I can’t claim durability. A knowledge system’s value depends on whether it survives a job change, a tool migration, and a few years of neglect. Mine hasn’t been tested against any of those.
The Real Filter
That’s the version of a second brain that makes sense to me. Not an enormous file telling an AI everything I’ve ever said or believed. A curated body of knowledge that helps both me and the model think better when the knowledge actually matters.
The goal isn’t to give AI more context. It’s to get better at deciding which context is worth keeping.
