The AI Note-Taking App That Organizes Itself: Vannevar Bush Called It in 1945

In July 1945, Vannevar Bush published an essay in The Atlantic Monthly about a desk that remembered everything he read. He called it the memex. It was never built. What he described is the thing every AI note-taking app that organizes itself is still trying to become.

Bush had spent the war coordinating American science. He watched thousands of researchers produce a record that outgrew anyone's ability to read it back, so he wrote down the diagnosis.

"Our ineptitude in getting at the record is largely caused by the artificiality of systems of indexing," he wrote. Store a document alphabetically and it lives in one place. To find it later you need rules about paths, and "the rules are cumbersome."

His alternative was associative indexing. The mind jumps from one thought to the next, he argued, so a machine should let you move the same way. Put two documents side by side, tap a key, and they are joined for good. Join several and you have a trail. Name it, and it can be recalled later, reviewed like pages, and copied into a friend's machine.

"The process of tying two items together is the important thing," Bush wrote. Storage was the easy half.

His own example is a study of the Turkish bow. Why did the short Turkish bow beat the English longbow in the Crusades? He pulls an encyclopedia entry, leaves it projected, finds a history that adds something, and ties the two.

The trail grows a branch into materials science once elasticity turns out to matter. He writes a page of his own analysis and joins it to the branch.

Years later, a conversation about why people resist innovation reminds him of the same trail. He taps a key, the head of it appears, and he runs the whole thing again. Then he photographs it and hands it to his friend.

Seven decades on, the personal information management research says we mostly still use folders. Ofer Bergman and Steve Whittaker found that search and tags, the methods information scientists prefer, work better for public information than for your own. People keep their own material in nested folders and remember roughly where it sits.

That is the honest verdict on the last 80 years. The trail was described in 1945 and the closest thing most of us built is a browser history.

What changed is who can do the tying.

The blocker in Bush's design was labor. Every link was a human decision made with a keyboard and a code book, at the moment the thought was fresh. People did not sustain it, and the machine he imagined did not exist anyway.

A model can now read what you captured and propose the trail. This note from March, the call transcript from June, the decision you recorded in August, all circling the same assumption, all in one line.

A trail is only as good as the record underneath it. Which means the capture has to happen while the thought is warm, months before anyone asks the question that the trail answers.

Compass is built for that half of the job: capture in seconds, then let the system propose the structure you would have drawn if you had the hours.

Bush's machine was never really about the desk. It was the habit of tying things together while the thought was still warm. Eighty years later, that habit is finally cheap.