What Darwin Was Afraid to Forget

There's a specific experience you get when you search your notes for evidence that you might be wrong about something. You find very little. Not because you're never wrong, but because that's not what you've been writing down.

In his autobiography, Darwin described his golden rule: "whenever a published fact, a new observation or thought came across me, which was opposed to my general results, to make a memorandum of it without fail and at once; for I had found by experience that such facts and thoughts were far more apt to escape from memory than favorable ones."

He wasn't describing a general note-taking practice. He was naming a specific failure mode in human cognition and responding to it structurally. We selectively encode information that fits our existing beliefs. Contradictions slip through. Darwin's solution was mechanical: capture the contradiction immediately, before memory could quietly dispose of it.

Why your knowledge base has a bias built in

You don't need to be unusually biased for this to happen. The problem is architectural.

When you read something that fits your mental model, you have language ready for it, context to connect it to, a slot to file it in. When you encounter something that doesn't fit, none of that scaffolding exists. The contradiction feels incomplete, unresolved. There's no obvious place to put it. So it doesn't get put anywhere.

Over years, a notes app becomes a record of your existing beliefs, annotated and elaborated. Not a knowledge base — a monument to your current priors. You can tell this is happening when searching your notes for counterarguments produces mostly silence.

This matters more than it seems, for two reasons. Time is the first: future-you has less context about how the notes were assembled and is more likely to take them at face value. AI is the second.

What an AI reads

When an AI reads your knowledge base, it inherits the skew. It finds the things you captured, which are already biased toward what you believed at the time you wrote them. It retrieves the note that confirms the conclusion you're already leaning toward. It doesn't know what you decided wasn't worth writing down.

We talk a lot about whether AI tools have biases. Less discussion goes to the data we hand them. An AI reading a knowledge base assembled without Darwin's rule will be confidently wrong in exactly the ways you are, except faster and with more apparent authority.

A language model reading your notes doesn't have a prior of "this person is probably a bit biased toward their own opinions." It reads the notes as a fair sample of what you know. When they're not, it can't tell.

What structure does for this

Darwin's solution was behavioral: write down the contradiction immediately. That's still good advice. But there's a structural piece underneath it.

Most note apps have one kind of object: the note. A note holds anything, which means it carries no metadata about what kind of thing it holds. Is this a belief you've examined carefully? A thing you read once and can't reconcile with something else you think you know? The note doesn't say, and you can't ask.

Typed structures help. A decision record that asks you to list alternatives considered is forcing a Darwin moment: name the options you rejected and why. A preference record with a "last reviewed" date makes staleness visible. An uncertainty field on a note is a declaration that something is still open.

None of this automatically captures contradictions. But it creates the vocabulary for them. And once there's a place to put a contradiction, you're more likely to notice when you've encountered one.

The note that escapes first

Darwin's observation — that contradictory facts escape memory faster than favorable ones — has since been confirmed repeatedly in cognitive research. We encode schema-consistent information more deeply. We forget what doesn't fit.

A knowledge base doesn't change this encoding process. It only inherits the output of it.

What changes is the opportunity for correction downstream. If the contradiction made it into a record, an agent reading the knowledge base can surface it later. It can notice that a decision was made on an assumption that a newer note quietly contradicts. It can flag a preference as potentially outdated. It can be honest in ways raw text retrieval can't manage. But only if the contradiction was written down in the first place.

Darwin kept notebooks full of things he was trying not to forget. What he was most afraid of forgetting was the evidence that might prove him wrong. That turned out to be the note worth keeping.

---

Asgeir Albretsen is the founder of Harbor.