Spaced Repetition for Technical Concepts

Spaced repetition has a simple core: review an idea shortly before you would otherwise forget it, and let the gaps between reviews grow as recall stays reliable. Applied to technical concepts, it can keep definitions and reasoning available for months without the cost of rereading everything. The method is easy to misuse, though, and most of the misuse comes from the cards rather than the schedule.

How the spacing works

When you recall something correctly after a delay, the memory is strengthened and usually stays available for longer the next time it is needed. When recall fails, the item should come back sooner. Scheduling programs follow this pattern: they lengthen the gap after each success and shorten it after a failure. The exact intervals differ between programs, and they are design choices rather than fixed laws, so treat any particular schedule as a starting point.

The practical consequence is that the review is the study. A card you glance at and immediately recognize has done little work. A card that asks you to produce an answer without seeing it is where the strengthening happens. Recall is effortful, and much of the benefit lies in that effort. Grading yourself is easy to do generously. A practical rule is to count a card as recalled only if you produced the key reasoning step without a hint, not if the answer looked familiar once you read it.

Write questions that test reasoning

Most technical cards fail because they ask for recognition of a phrase. "What is a dot product?" invites a memorized sentence. Ask instead for a use or a consequence: "Why does a dot product near zero between two feature vectors mean they share no direction in that comparison?" The answer has to be built, not recalled.

Good reasoning cards can be short. "Why does a softmax output stay the same if the same constant is added to every score?" is a good card. The answer is that softmax depends only on differences between the scores, so the added constant cancels when the exponentials are divided by their total. Someone who has to produce that explanation is more likely to understand softmax than someone who can recite its formula. Another good card asks why a smaller learning rate might help once a loss curve starts oscillating, since the answer requires connecting the step size to the overshoot. A card on a working habit can work the same way. Ask what the first thing to check is when a model's output changes after a prompt edit that looked harmless, and why. The answer is a sentence about isolating the change, which a recognition card would never test.

For procedures, a card can ask for the steps in order and for the reason one step depends on the one before it. For each card, write the answer as one or two sentences in your own words. If you cannot write it, the card is too large, or the concept needs a worked example before it can be reviewed.

A weekly review habit

Set aside about fifteen minutes each week for a review pass. Go through the cards that came due, and mark each as recalled fully, recalled with hesitation, or missed. For cards marked hesitant or missed, rewrite the question so it asks for the reasoning more directly. For cards recalled easily but not explained, add a second question that requires application.

Delete cards that are wrong or no longer useful. Old cards can become a trap, because they keep returning and take time from the material that needs work. A small deck reviewed every week is worth more than a large deck that is always behind. Adding cards faster than they can be reviewed is the most common failure of the method. Keep new cards to a number the weekly review can absorb without growing past the time available.

The limits of memorization

Spaced repetition strengthens what you can retrieve. It does not replace understanding, and it does not tell you whether the answer you recalled is correct. If a card was written with an error, the schedule will faithfully preserve the error. Check new cards against a primary source, such as a book, the documentation or your own worked example, before they enter the rotation.

The method also does little for skills that require a sequence of actions in a changing context, such as debugging a system you have not seen before. Cards can support those skills by holding the concepts they depend on, but practice with real problems is still necessary. Treat cards as scaffolding for reasoning, not as a substitute for it. A deck can also grow around trivia that is easy to test and unimportant to the work. If a card would not change what you do in practice, it is a candidate for deletion.

Try a short cycle

For the next two weeks, write ten cards from the material you are studying now. Each card should ask for an explanation in a sentence, not a single word. Review them on whatever schedule your tool uses, or simply after one day, three days and one week. Keep a note of the cards you missed. Where a card comes from a source note in your notes system, link the two, so that a failed card leads straight to the explanation.

At the end of the cycle, rewrite the three cards you failed most often, and read the reasoning behind their answers before trying them again. Then check which cards you could use to explain the idea to a colleague without looking. Those are the cards doing their job. The rest need rewriting rather than more repetitions.