Connecting the dots: how links make memory, and when they don’t

A paragraph about laundry

Here is a passage from a well-known experiment. Read it once.

The procedure is actually quite simple. First you arrange things into different groups. Of course, one pile may be sufficient depending on how much there is to do. If you have to go somewhere else due to lack of facilities that is the next step; otherwise you are pretty well set. It is important not to overdo things. That is, it is better to do too few things at once than too many.

It reads like nonsense — grammatical, but hard to hold on to. Now read it again knowing that it is about washing clothes. Every sentence clicks into place.

John Bransford and Marcia Johnson used passages like this in 1972. People who were told the topic before reading rated the passage as far more comprehensible and recalled much more of it than people who weren’t told. People told the topic after reading did no better than people never told at all (Bransford & Johnson, 1972). The words were identical. What changed was whether readers had something to connect them to while they read.

Schemas, and what you already know

Psychologists call the structures you understand new things through schemas: organised knowledge about how some part of the world works. Your laundry schema includes sorting, machines, laundrettes and overloading. Once it is switched on, the passage stops being a list of vague instructions and becomes an account of something you know.

The educational psychologist David Ausubel built a theory of learning on this. In 1968 he wrote that the most important single factor influencing learning is what the learner already knows — find that out, and teach accordingly. New material is learned well when it is anchored to something already understood, and poorly when it floats free.

Two practical consequences follow:

• Before learning something new, say what you already know about it. Even a few lines switches on the right schema and shows you where the new idea will attach.
• After learning something new, connect it to something else you know. An idea with no connections is an idea you will lose, and a connection made now is a retrieval route later.

What concept maps do

If connections matter, why not draw them? Concept maps — diagrams with ideas as nodes and labelled lines between them — have been studied for decades. A meta-analysis by John Nesbit and Olusola Adesope covering 55 studies found that working with concept maps was associated with better knowledge retention than reading text, attending lectures or discussing (Nesbit & Adesope, 2006).

So maps help. But the more interesting question is which part of working with a map helps — looking at the connections, or making them.

The surprise

In 2011 Jeffrey Karpicke and Janell Blunt ran a study that should be read by anyone who draws diagrams of their notes (Karpicke & Blunt, 2011). Students read a science text and then either built a detailed concept map of it with the text in front of them, or practised retrieval: they put the text away and wrote down everything they could remember, then read it again and did it once more.

A week later, the retrieval group did better — on questions about details and on questions that required inference. They did better even when the final test was to draw a concept map, the very task the other group had practised. And the students themselves predicted the opposite: they expected concept mapping to work better.

The lesson isn’t that maps are useless. It is that the learning comes from pulling connections out of your own memory, not from seeing them laid out. A beautiful graph of your notes is pleasant to look at. On its own, it teaches you very little. Stating a connection from memory — this matters to that, because… — is the part that works.

Generative links: always say why

This suggests a simple rule for anyone keeping linked notes: a link should carry a reason, written by you.

Compare two links from a note on attention to a note on embeddings:

• See also: [[Embeddings]]
• [[Embeddings]] — queries, keys and values are projections of the embeddings, so attention can only compare what the embeddings already encode.

The first is a pointer. It costs nothing to make and does nothing for your understanding. The second required you to retrieve both ideas, work out how they relate, and commit to a claim. It may be wrong — which is useful, because a wrong link written down is a link that can be checked. A wrong link you only felt is invisible.

The same rule helps when you read someone else’s notes, or an AI’s explanation, that asserts a connection. Before accepting it, try stating the reason yourself. If you can’t, you haven’t understood the connection; you’ve been told about it.

This is also why the obvious connections and the meaningful ones deserve different treatment. Linking every mention of Attention to the note called Attention is bookkeeping, and can safely be automatic. Saying how attention relates to the cost of long documents is learning, and shouldn’t be.

A daily practice: three connections

Here is a practice that takes ten minutes and uses all of the above.

1. Pick three things you learned or noted recently. They can be from the same subject or from different ones.
2. For each, close your notes and ask: what does this connect to? Look for one thing you already understood well.
3. Write the connection as a sentence with a because in it. Inflation expectations matter to wage negotiations because workers ask for raises based on the prices they expect, not the prices they’ve seen.
4. Then check it. Open both notes. Does what you wrote agree with them? If it doesn’t, you’ve found something worth a question.
5. Keep the sentence on one of the two notes, as the link between them.

Three a day is enough. After a month you will have around ninety connections, each one stated from memory, and a notebook whose structure reflects what you understand rather than how you happened to file things.

How Graphite does this

Graphite treats connections the way this research suggests. A [[link]] you write shows on both notes, and a link to a note that doesn’t exist yet stays dimmed, as the next thing to write. For the meaningful connections, the Guide asks rather than tells: Link related proposes the links a note should have, each with its reason, and adds none until you accept — and the link that counts is the sentence you write about why. The Map shows the shape that results, but the learning happens in the sentences. You can try it in your browser.