Part 2 of a three-part series. Part 1: Personal Wikis and the Desire to be Seen → · View the Cortex project →
In Part 1, I argued that the real pull of Karpathy’s LLM Wiki wasn’t the architecture — it was an emotional promise: let the machine see you, and you’ll finally see yourself. A personal wikipedia that writes itself, maintained for free, compounding forever.
Good pitch. I wanted to know if it was true.
So I built one. Not a fork of Karpathy’s gist or Farzapedia or any of the implementations that bloomed in the tweet’s wake. I wanted to start from a more fundamental question: what would a knowledge system look like if it was designed around how memory actually works — not just how files get organized?
I call it Cortex. What follows is what I learned building it, what surprised me, and where the promises from Part 1 held up — and where they quietly fell apart.
The Starting Point
I had 660 links saved to Notion. Articles, papers, GitHub repos, blog posts — things I’d read, half-read, or bookmarked intending to read. Years of digital pocket litter.
The first version of Cortex was straightforward: parse the content, store it in a database, render a list. A wiki of your links. It worked. It was also completely uninteresting, because a scrollable list of 660 items is not self-knowledge — it’s just a longer Notion.
The interesting question was: what should happen when you look at all of them at once?
The Shape of a Mind
The original attempt at connecting everything produced what you’d expect: a hairball. Every article connected to every other article through vague keyword similarity. Dense in the middle, incomprehensible. It said nothing about me that a word cloud wouldn’t.
The breakthrough was a shift in framing. Stop connecting articles to each other. Instead, extract topics — not from my tags, which reflected what I thought I was reading about, but from the content itself — and let articles cluster around those topics naturally. Topics became large, labeled nodes. Articles became the texture around them.
The resulting graph looked like a solar system. Topic hubs orbiting a center, articles clustered around them like dust. And immediately, it showed me something I hadn’t expected.
I thought I read broadly about AI. I would have told you that if you’d asked. The graph showed I read almost exclusively about AI infrastructure and tooling — very little theory, almost no alignment, nothing on ethics. That gap between what you think your interests are and what your actual reading behavior shows is uncomfortable in a way that no amount of journaling has ever produced.
That was the first moment the promise from Part 1 felt real.
The Invisible Failure
The second major piece was a timeline: show me how my reading patterns have changed over time. Monthly activity, category distribution, the shape of curiosity across months and years.
The timeline showed nothing. Blank. “Not enough data yet.”
There was data. 660 links. The problem was that every single date had been silently discarded during import. Not because the dates were missing — they were right there in the source files, perfectly formatted. The parsing code read them, tried to interpret them, failed, and instead of complaining, quietly wrote nothing. Six hundred and sixty records, all successfully imported, all silently lobotomized of their most important dimension.
The root cause was a single default setting. The date parser didn’t know what language to expect. It saw “March” and “PM” and — on a system not configured for English — returned nothing. The data was perfect. The logic was correct. The infrastructure worked. And the entire timeline was invisible because of an assumption so small that no reasonable person would have thought to question it.
One line fixed it. One property, added in one place, and all 660 dates materialized.
I keep thinking about that bug. Not because it was technically interesting — it wasn’t — but because it’s a near-perfect metaphor for the thing this essay is trying to explore. How much of what you think you know about yourself has been silently discarded by defaults you never examined? How many of your own timestamps are NULL — not because the experiences didn’t happen, but because you never took the time and space to properly reflect and interpret? We’re all in a mad day to day rush through life. The ancient wisdom of stop and smell the roses, meditate, pause, take a moment for gratitude is so relevant here it’s honestly infuriating. Like a mother whose words echo in your head decades later.
Memory Made Visible
Once the dates came back, the timeline came to life.
I could see the month I discovered a new field — a sudden cluster of articles on a topic that hadn’t existed in my reading before. I could see the month I stopped saving anything at all, which correlated with a period of burnout I’d half-forgotten. I could see the seasonality of my curiosity: more reading in winter, less in summer. Obvious in retrospect, invisible until rendered.
The strangest part was the gaps. Months where I’d saved nothing, or saved only one thing. Not because I wasn’t consuming information — I was, constantly — but because I wasn’t capturing it. The timeline didn’t show my reading. It showed my intention to remember. And those are very different things.
Looking at Cortex’s timeline felt less like reviewing a record and more like reading a biography written by someone who’d been watching me from across the room. Accurate, but distant. Observant, but missing the interior monologue.
The Promise, Half-Kept
In Part 1, I described the implicit bargain of the personal wiki: the longer you use this system, the smarter and more attuned to you an AI will become. You feed it your information, and it shows you the shape of your own mind.
After building Cortex, here’s what I can tell you: the promise is half right.
The machine does see you — but it sees your behavior, not your intent. It sees what you saved, not why you saved it. It sees how often you return to a topic, but not whether you actually thought about it. The gap between saved and absorbed, between bookmarked and internalized, is not something a knowledge graph can close.
I expected to feel seen. What I actually felt was audited.
Not in a hostile way. More like the difference between a close friend and a good analyst. The friend says, “you’ve been really into infrastructure lately — what’s driving that?” The analyst says, “your infrastructure content represents 47% of total saves, up from 31% last quarter.” Both are true. Only one feels like being known.
Karpathy’s original intuition — that a system like this could compound your knowledge — holds more for the retrieval half than the growth half. Cortex makes it easier to find what you’ve already encountered. It does not make you encounter better things. It does not tell you what you should read next, what gaps exist in your thinking, or whether any of this is making you wiser. That work is still yours to do.
And the uncomfortable question from Part 1 — when an AI writes Wikipedia about you, is the result self-knowledge or self-flattery? — has an answer now: it’s neither. It’s a behavioral audit. The machine is a mirror, but it reflects your actions, not your face. Whether you like what you see is up to you.
What It Means
The system categorizes with confidence regardless of whether it’s right. An article about AI ethics might be filed under AI Fundamentals rather than Philosophy, because the system responds to keywords, not meaning. It captures the shape of your attention and misses the substance of your engagement.
This is the tension I didn’t anticipate. In Writing as Proof of Human, I argued that the costly signal of human writing is the residue of reasoning — the fingerprint of unique thought. A personal wiki inverts that: it’s a fingerprint of unique attention. What you saved, when you saved it, how it clusters. That’s real, and it’s revealing. But attention is not the same as understanding. And a map of what caught your eye is not the same as a map of what changed your mind.
In Most People Stop at Step One, I built a tool to practice tracing your work upward — from task to feature to capability to strategic consequence. Four hops. The interesting thing about Cortex is that it does the opposite: it traces your interests downward, compressing years of curiosity into clusters and trends. It’s powerful. But it stops at description. It doesn’t ask the “so what?” It shows you the solar system of your attention and leaves the interpretation to you.
Which is, I think, the honest answer to whether the promise was fulfilled. The machine saw me. It showed me patterns I couldn’t have seen alone. But the work of understanding what those patterns mean, and deciding what to do about them — that’s still mine. There is no escalator to self-knowledge. There’s a better mirror and a well-lit room. You still have to look.
The Question That Won’t Let Go
Here’s what I didn’t expect to be thinking about after building Cortex, and what I can’t stop thinking about now.
If a tool like this can surface my behavioral fingerprint — my actual interests, my working patterns, the shape of my attention over time — then it can do the same for anyone. And if everyone has a legible, machine-readable summary of how they think and what they care about, the implications go well beyond personal productivity.
The question stops being can the machine see me? and starts being what happens when everyone is visible?
That’s Part 3.
Cortex is available on GitHub. It is an open-source macOS app built in Swift. The build process, including every bug and breakthrough described above, was done in collaboration with Claude.