What Your Notes Can't Tell You
Every note looks equally authoritative when you read it later. Most of them weren't.
I have a note from 2022 that says a colleague prefers working asynchronously. For a couple of years, I assumed that was just true. Then I found out she'd actually been hoping for more real-time collaboration.
The note was probably accurate in some limited sense. I'd had a conversation, formed an impression, written it down. But what I'd written was a working hypothesis, and I'd stored it as a confirmed fact. The note had no way to tell the difference. Neither did I, reading it back.
This isn't a memory problem. Memory failed too, but that's expected. The problem is that the note itself provided no signal about how confident I was when I wrote it. It looked exactly as authoritative as a note where I actually knew what I was talking about.
What you capture when you write something down
In 2002, psychologists Leonid Rozenblit and Frank Keil at Yale ran a study where they asked people to rate their understanding of everyday objects — how a zipper works, how a toilet flushes. People rated their understanding reasonably high. Then they were asked to explain the mechanism in detail. Ratings dropped sharply, often by half. The effect became known as the illusion of explanatory depth: people feel they understand things with greater depth than they actually do, specifically for causal and mechanical knowledge.
Notes are written in the moment before that test. You write something down while it feels understood — while the mental model is fresh and the confidence is high. The note doesn't survive Rozenblit and Keil's experiment, because it's never asked to explain itself. It just sits there, emitting whatever authority the original feeling of understanding provided.
Two years later, reading it back, that feeling is long gone. What's left is the text. And the text looks the same whether you were certain or guessing.
A convention that solved this for public writing
There's a practice, common on LessWrong and the EA Forum, of prepending posts with an "epistemic status" — a sentence at the top that signals how much confidence the author is actually staking on the piece. "Epistemic status: speculative, mostly thinking out loud." Or: "Fairly confident, based on direct experience over several years." Or: "I removed the hedges for readability but don't take this as settled."
The convention exists because readers have an information asymmetry. They can't know whether the author spent six months on a question or dashed something off in an hour. The epistemic status header closes that gap — not by certifying truth, but by signaling what kind of claim this is.
Future-you reading your old notes has exactly the same asymmetry. You don't know whether the person who wrote it was certain or guessing. You don't know if they heard it directly or inferred it. You don't know whether "prefers async" was said explicitly or read between the lines of one conversation after a rough week.
Nobody puts epistemic status headers on personal notes. Of course they don't — the form doesn't invite it, and in the moment of writing you're not thinking about the stranger who'll encounter it later. But that stranger is you.
What structure does instead
A note that says "Marta prefers asynchronous communication" is a different epistemological object from a structured record that says: person named Marta, preference for asynchronous communication, sourced from one conversation in March 2022, marked tentative.
These contain the same facts. But filling in a schema requires a kind of epistemic decision that prose doesn't. You have to choose: is this a preference or a habit? Is it something she said explicitly or something I inferred? Is the source one data point or a pattern across several interactions?
Not because the form asks those questions directly. Because choosing what type of thing you're writing — preference, not note; tentative, not confirmed — forces you to locate the claim in a space of possible claims. The label carries signal that flat prose can't.
This is partial. A schema doesn't prevent you from entering speculative things with false certainty. You can confidently type something wrong into a well-designed form. But when you read it back, the type still tells you what kind of assertion was being made. A preference is labeled a preference. A decision is labeled a decision. The difference between "I think" and "I know" is at least recoverable — you have to actively mislabel the claim, rather than having it disappear into undifferentiated prose.
The asymmetry you can't fully close
There's probably no complete solution here. People won't annotate personal notes with confidence intervals. Most of the time they shouldn't have to — a note is a note, not a research paper. And the underlying issue, that writing something feels like understanding it, is older than any software that might help.
But you can build a system where the difference between "I confirmed this" and "I wrote this down once" is at least not erased on contact. Where the structure of what you store carries some trace of what kind of claim it was. Where, when something turns out to be wrong, you can see that you marked it tentative and chose not to.
The note about my colleague is still there. I've updated it. I've also started writing notes a little differently — more carefully about what I'm actually asserting and what I'm just recording as an impression. It's less about the software and more about the practice. Though the two are not as separable as they sound.
Asgeir Albretsen is the founder of Harbor.