Spaced Repetition, With and Without AI

Spaced repetition — reviewing material at increasing intervals, timed to just before you would forget it — has better evidence behind it than any other study technique. It predates computers and the software versions have been good for two decades.

AI changes the cost of producing cards. It does not change the method, and it introduces a specific way to waste your time.

Why it works

Retrieval strengthens memory more than review does. Trying to recall something, and struggling, does more than re-reading it. This is the single most robust finding in the learning literature and it contradicts how most people study.

Difficulty is the point. A card you recall instantly does little. A card you almost fail does most of the work. Software that schedules for roughly this difficulty is doing the important thing.

Spacing beats massing. The same total time distributed over weeks produces far better retention than the same time in one session.

What makes a good card

This is where most people fail, and where generated cards fail hardest.

One fact per card. A card asking three things cannot be scheduled sensibly, because you will know one and not the others.

Tests recall, not recognition. "What is X" beats a multiple choice. The struggle is the mechanism.

Short. If it takes twenty seconds to read, it is not a flashcard.

Unambiguous. A card with several defensible answers produces frustration and unreliable scheduling.

Meaningful to you. Cards made from your own understanding, in your own words, are recalled better than imported ones. This is the same processing effect that makes note-taking work.

Where AI helps

Mechanical cards. Vocabulary, dates, definitions, formulae, terminology. Generating these from a document is fast and the quality is adequate because the content is simple.

Reformatting. Turning your own notes into card format.

Generating variations to test the same fact from different directions, which strengthens the association.

Explaining a card you keep failing. The most useful application: when you repeatedly get something wrong, asking why it works that way frequently fixes it permanently.

Suggesting what to make cards about from a chapter.

Where it hurts

Bulk generation without review. Two hundred generated cards from a textbook chapter will include duplicates, ambiguous questions, trivia, and cards testing recognition. You then spend months reviewing bad cards.

Skipping the making. Writing the card is part of the learning. Generating cards for the difficult conceptual material removes exactly the processing that would have helped.

Cards that are too long. Models produce complete, well-formed sentences, which is wrong for this format.

A workable split

Write your own cards for concepts — anything where making the card requires you to decide what the essential point is.

Generate cards for the mechanical layer — terms, dates, vocabulary, syntax.

Review every generated card before it enters the deck. Delete the ambiguous ones, split the compound ones, shorten everything. This takes a fraction of the time it takes to review a bad card fifty times.

Delete aggressively. A card you have failed ten times is a bad card or material you do not understand. Either rewrite it or go and learn the underlying thing.

Practical operation

Review daily, briefly. The schedule assumes it.

Do not skip days and then binge. The intervals stop meaning anything.

Keep the deck small enough to sustain. A deck you abandon teaches you nothing. Two hundred good cards reviewed daily beats two thousand reviewed never.

Add as you learn, not in bulk afterwards.

Accept that the first weeks feel slow. The benefit is retention at six months, which is invisible at week two and is the entire point.

Fixing a card you keep failing

Repeated failure on the same card is information, and the usual response — reviewing it more — is the wrong one.

First ask whether the card is bad. Ambiguous, compound, or testing something that only makes sense in context. Most repeatedly failed cards are badly written rather than hard.

Split compound cards. A card asking two things fails whenever you know one.

Add a cue. Not the answer — a hint that narrows the retrieval without giving it away.

Check you understand the underlying thing. Failing a card repeatedly frequently means you memorised a fact without the concept behind it. Go and learn the concept; the card becomes easy.

Reformulate rather than repeat. Rewrite the card from scratch in your own words, which usually fixes it.

Delete it if it does not matter. A card you have failed twenty times, on something peripheral, is costing you months of review for nothing.