The conversation in the margins

On September 28, 1838, Charles Darwin read a passage in Malthus's Essay on the Principle of Population about the pressure of scarce resources on competing populations. In his notebook — Notebook D, the third in his transmutation series — he wrote a response that would take him twenty more years to fully develop. The note isn't a standalone thought. It lives in the margin of that moment: Darwin talking back to Malthus, extending an argument, testing it against what he already knew about breeding and variation. You can't separate the note from what it was responding to.

That's what marginalia actually is. Not a note. A reply.

What the margin preserves

H. J. Jackson, who spent years studying annotated books in libraries across England, described Samuel Taylor Coleridge's marginalia as eight thousand notes, "varying from a single word to substantial essays." Coleridge treated books like correspondence. His notes on Swedenborg, on Kant, on Shakespeare read as half of a dialogue. They only make sense in context. They were meant to.

This isn't accidental. The physical format enforces it. When you write in the margin of a book, the source is right there. Your thought sits next to what prompted it. The relationship is structurally preserved, not reconstructed from memory. Which means three years later, when you open that book again, you can have the conversation again. You know what your note was responding to.

Modern tools have quietly dismantled this.

The extraction problem

The contemporary workflow goes like this: highlight a passage in Kindle. It syncs to Readwise. Every morning, you review fragments. The good ones go into Obsidian or Notion, tagged and filed, often with a sentence of your own thinking attached. The original passage may or may not come along. The surrounding paragraphs almost certainly don't. By the time a note lands in your knowledge base, it's been stripped three times — from the paragraph, from the chapter, from the book. You have the response but not the prompt. The reply with no letter.

The result is a specific kind of decay that feels invisible until you go back. I've returned to notes I wrote with real conviction — "this changes how I think about X" — and had no idea what they were responding to. The note seemed important. I couldn't reconstruct why.

This isn't a flaw in Readwise. It's structural to any system that treats notes and sources as separate artifacts. The extraction is the problem.

What AI reads when it reads your notes

The extraction problem gets considerably worse when AI enters the picture.

When you re-read an old note, you at least have context for your own thinking — your interests at the time, what you were working on, the rough period when the idea arrived. You can reconstruct some of the conversation even when the source is gone. An AI reading your notes has none of that. It reads them as statements. As facts. As things-you-believe, not things-you-thought-while-engaging-with-something-specific.

A margin note in a copy of Malthus is contextually rich. You know the author being read, the passage being responded to, the relationship between the reader's thought and the source. A decontextualized note in a knowledge base is much poorer: a claim floating free of its origin.

A note that says "population pressure creates selection" could be Darwin's synthesis of Malthus, your half-formed interpretation of a podcast, or a thought you had in the shower on a Tuesday. The text doesn't say. In the margin, the answer is obvious. In a flat document, it isn't — and the AI treating it as an authoritative statement is making a kind of structural error.

What structure does instead

Typed entities survive this better than prose notes, and the reason is simpler than it sounds. A person record doesn't decay contextually — it just is. It says: here is Marcus, here is when we last spoke, here is what he works on. The facts carry their meaning regardless of where they came from. A preference, a task, a decision with its context recorded — these are translated from the conversation into something self-contained.

The prose notes — the ones that look like thinking rather than recording — are the ones that keep their dependence on context. Those are the orphaned replies. That's where the decay lives.

I don't think there's a clean solution. Linking every note back to a source helps, but sources go stale, move, disappear. The real answer might be something like what Darwin actually did: write in the source itself. Keep the response next to the prompt. Let the conversation stay whole.

But most of us have stopped writing in our books.

---

Asgeir Albretsen is the founder of Harbor.