Ideas are not folders

Where does Attention go?

Suppose you are learning how large language models work, and you have just written a good note on attention — the mechanism that lets each word in a sentence weigh every other word when working out what it means. You go to save it. Where does it go?

The obvious answer is a folder called Transformers. But next month you start reading about image models, and attention turns up there too, doing the same job on patches of a picture. Then you read about how people pay attention, and the word means something related but different. The note now belongs in three places. A folder lets it live in one.

Most people solve this the way the folder forces them to: they pick a home, and the other two connections quietly disappear. Or they copy the note, and now there are two versions drifting apart. Or they give up on folders and dump everything in one place, which works until there are five hundred notes and none of them can be found.

The problem isn’t that people are bad at folders. It’s that folders encode a claim — every thought has exactly one right place — that is false for understanding.

Understanding is many-to-many

What you know isn’t a tree. A single concept usually helps explain several bigger ideas, and a big idea always rests on several concepts. Marginal cost belongs to how prices are set, to why software is cheap to copy, and to why airlines sell the last seats on a flight for less. Remove any one of those connections and you understand each idea a little worse.

A folder tree is a good shape for things that really do have one home: invoices by year, photographs by trip. It is a poor shape for the thing a learner is trying to build, which is a web of ideas where the interesting part is how they relate.

So the question isn’t which folder? It is what is this note about, and what does it connect to? — and the answer is usually more than one thing.

Big ideas and essential questions

Teachers have a useful vocabulary for this. In Understanding by Design, Grant Wiggins and Jay McTighe argue that good courses are organised around a small number of big ideas and essential questions, rather than a sequence of topics (Wiggins & McTighe). A big idea is something worth understanding in its own right, that many smaller facts and skills hang from: how transformers work, why prices rise, how evolution produces complex organs. They list six facets of real understanding, and the first is the ability to explain.

That gives a learner a better unit than the folder: a big question you are trying to answer, which works as an anchor. It doesn’t contain notes the way a folder contains files. Notes point to it. One concept can point to several big questions, and each question can gather everything that points to it — its concepts, your own thoughts, the smaller questions you haven’t answered yet.

The test of whether something is a big question is simple: could you spend an afternoon on it and still have more to learn? Attention is a concept. How do transformers work? is a big question.

An idea is a kind of note

If the big question is the anchor, what is an idea? It is one of the things you write down on the way: a thought of your own — a hunch, an analogy, a what-if — as opposed to a fact you can check, a claim somebody argues, an example, someone’s exact words, or the source they came from.

Discourse graphs, a research tool for synthesizing what you read, are built from just such kinds: questions, claims, evidence and sources (Chan et al., 2024). A note that knows it is a claim reminds you to look for who disputes it; a fact reminds you to say who found it, and when. Keep the kinds few and fixed, and let tags carry everything else: people resist structure they have to commit to before they know what they think (Shipman & Marshall, 1999).

Associative, not hierarchical

People who have thought hard about notes tend to arrive at the same place.

The sociologist Niklas Luhmann kept a famous slip-box of index cards for decades, numbered so that a new card could be filed next to the one it continued, and cross-referenced by hand to any other card it related to. The structure that mattered was the web of references, not the order of the drawer.

Andy Matuschak’s writing on evergreen notes puts it most directly: notes should be atomic, concept-oriented and densely linked, and you should prefer associative structure to hierarchical taxonomies (Matuschak). A taxonomy asks you to decide in advance how the world is divided. A web of links lets the structure emerge from what you actually understand.

Two practical rules fall out of this, and both are worth adopting whatever tool you use:

• One thought per note. A note that makes two claims can’t be linked precisely; you never know which of the two the link means.
• Title notes as claims. Attention is recomputed for every input tells you what the note says before you open it. Attention thoughts tells you nothing, and invites a second, vaguer note next to it.

The collector’s fallacy

There is a trap here, and it is worth naming because organising notes makes it worse.

Christian Tietze calls it the collector’s fallacy: “to know about something isn’t the same as knowing something.” Saving an article, highlighting a passage, filing a note in exactly the right folder — each of these feels like progress. None of them is understanding. You can spend an evening reorganising a notebook and come out knowing nothing you didn’t know at the start.

Filing is the most seductive version, because it looks like thinking. Deciding whether a note goes under Economics or Policy feels like an intellectual act. It mostly isn’t. The intellectual act is deciding how two ideas relate and saying it in a sentence — which is something no folder asks you to do.

So it is worth separating the two jobs:

• Filing — putting each note where it belongs, keeping track of what is about what — is clerical. It should take as little of your attention as possible.
• Connecting — saying why this idea matters to that one — is learning. It should take as much as it needs.

So: the notebook files, and you connect

If filing is clerical, a notebook can do much of it for you, as long as it follows some firm rules. It should tell you every time it files something and why, and let you undo it. It should never change your words. It should never delete anything, and never move a file you put somewhere yourself.

And it should leave the meaningful connections to you. A notebook can safely link an exact mention of a concept to that concept’s note — that is just bookkeeping. But when two ideas you’ve learned belong together for a reason, the useful thing for it to do is ask you what the reason is, and keep your answer as the link. Retrieving and stating a connection is what builds understanding; being shown one mostly isn’t (Karpicke & Blunt, 2011).

What it looks like on disk

Whatever tool you use, you can keep this structure in plain text files, so it outlives the tool. A note can say what kind of thought it is, and which big questions it belongs to, in its frontmatter, as links:

---
kind: idea
explainers: "[[How transformers work]]"
---

The membership lives on the member. The big question’s own note doesn’t keep a list of what belongs to it that has to be updated by hand; it finds them by looking at what points to it, the way backlinks work in Obsidian. A note that belongs to two questions simply names both. Nothing goes stale, and nothing needs copying.

Folders can still exist — Notes, Concepts, Explainers — but only as a place for each kind of file to live, never as a claim about what anything is about.

How Graphite does this

Graphite keeps the two jobs apart. A folder or a project is where a note lives while you learn about something or make something of it — a note is in one at a time, said by one line of plain project: or topic: frontmatter, so its file never moves. Links say how ideas connect, as many as a note needs, each with its reason: Link related suggests them and adds none until you accept. Once you turn on File my Inbox in Settings → AI, the Guide files your Inbox into the project or folder each note clearly belongs to. Every act is in the Logbook, with Undo; it never changes your words, never deletes, and never moves a file you put somewhere yourself — a photo or a PDF in the Inbox moves once, when it’s filed as a receipt. A note you keep in a folder of your own on disk stays there, and the Guide leaves it out of filing. If you already keep notes in Obsidian, the files stay where they are and work in both. You can try it in your browser.