Spaced repetition, actually applied.

The problem is not that you forget. It is when.

Forgetting is not a defect in the way memory works — it is most of how memory works. Anything you learn once and never use decays, and it decays fastest in the first day or two. Read a chapter on Monday, do nothing with it, and by Friday you have kept the shape of it and lost the substance: you know the chapter discussed three mechanisms and you could not name the second one.

The useful consequence is that decay is predictable. It is not random which facts you lose, and it is not random when. That predictability is what makes scheduling possible, and scheduling is the whole technique.

Review a thing just before you would have forgotten it, and the next interval you can leave is longer than the last one.

That is spaced repetition in one sentence. Every complication after this is about estimating "just before".

Two mechanisms, not one

It is worth separating the two effects, because most people apply one and assume they have both.

Spacing. The same total study time distributed across days produces markedly better retention than the same time in one block. Four twenty-minute sessions beat one eighty-minute session — not slightly, and not only at the end. This is the finding that makes cramming a strategy for Thursday's exam rather than for knowing the subject.

Retrieval. Being asked a question and having to produce the answer strengthens the memory far more than reading the answer again. This is the effect people skip. If your "review" consists of re-reading your notes, you get the spacing benefit and almost none of the retrieval benefit — and the retrieval benefit is the larger of the two.

Failing a retrieval attempt is still useful. Trying to remember and coming up empty, and then being told the answer, beats being told the answer without trying. This is deeply counterintuitive and reliably true, and it is the reason a study method should feel harder than re-reading does.

The intervals

You can do this by hand with index cards and a set of boxes — the Leitner system — and for a hundred facts it works. Past that, the bookkeeping is the thing that kills it, because each fact is on its own schedule.

The standard algorithm is SM-2, and it is worth knowing what it actually does because it is what most software including this one implements:

1. First successful recall: come back in 1 day.
2. Second: come back in 6 days.
3. After that, the next interval is the previous interval multiplied by an ease factor that starts at 2.5. So 6 days becomes 15, then 38, then 95, then eight months, and it keeps stretching.
4. The ease factor moves with how well you recalled it. Effortless recall nudges it up so the item comes back less often; a correct-but-difficult recall nudges it down. It never falls below 1.3, which is the floor that stops a hard item collapsing into a daily nag.
5. A failure resets the interval to 1 day and lowers the ease. The item has to climb the ladder again, and it climbs it more slowly than the first time.

The precise numbers are less important than the shape: intervals grow multiplicatively for things you know, and collapse to one day for things you do not. Anything with that shape works. A fixed schedule — "review everything weekly" — does not, because it spends the same effort on the fact you have known for a year as on the one you failed yesterday.

Where flashcards stop being the right tool

Spaced repetition software is nearly always flashcard software, and that conflation causes real problems. Spacing is a property of a schedule; a flashcard is one format. They are not the same thing, and the format has a narrow range.

Flashcards are excellent for:

• Vocabulary, in either direction.
• Definitions, formulae, constants, dosages, dates.
• Pairs — a term and its meaning, a symptom and its cause.
• Discriminations, if you write them as a pair on one card rather than two cards that never appear together.

They are poor for:

• Procedures. A card that asks "what are the seven steps" tests whether you can list them, which is not the same as being able to do them in order under time pressure. Space the practice instead: perform the procedure again after a day, then a week.
• Understanding a system. If you can answer every card about a mechanism and still cannot predict what happens when you change one input, the cards measured the wrong thing.
• Anything requiring judgement. "Which approach fits this case" has no back of the card. It needs a worked scenario and someone or something willing to say your answer was wrong and why.

The general fix is to space the right kind of retrieval rather than forcing everything into a question-and-answer pair. A returning procedure should return as a procedure to carry out; a returning distinction should return as two cases to tell apart.

Why almost nobody keeps it up

The technique has been well established for decades and the tools are free. The failure mode is nearly always the same, and it is not motivational.

You use it well for three weeks. Then you get ill, or a deadline lands, and you miss ten days. You open the app to 400 overdue cards, all of them "due", most of them things you actually still know. Clearing that pile would take two hours you do not have, so you do not open it tomorrow either, and the pile grows. Within a fortnight the tool is uninstalled — and the story people tell afterwards is "spaced repetition did not work for me", when what failed was the queue design.

Two things fix it, and both are decisions the software has to make on your behalf:

• A backlog must not accumulate as a debt. Work that was due during a fortnight you were absent should re-flow into the days you actually have, at the rate you can sustain — not appear as a single overdue heap.
• Review has to be part of the plan, not a second app. If today's reviews and today's new material are in different places, the reviews are what get skipped, every time.

How to start, if you are doing this by hand

1. Pick one subject. Applying this across four subjects at once is how the daily load becomes unmanageable in week two.
2. Write items only for things you have already understood. Spaced repetition maintains knowledge; it does not create it, and drilling a fact you never understood produces a memorised sentence and no ability.
3. Keep each item to one fact. An item asking three things fails as a unit and teaches you nothing about which third you missed.
4. Grade yourself honestly. The algorithm is driven entirely by your own report of how hard recall was, so charitable grading produces a schedule that quietly stops working.
5. Set a daily cap you would keep on a bad day — ten minutes, not forty. The schedule survives a small daily commitment and does not survive an ambitious one.

What BiliMind does with this

Since this is on BiliMind's site, the honest disclosure: the review system here is SM-2, with the intervals described above — 1 day, 6 days, then interval × ease, ease starting at 2.5 and floored at 1.3, a lapse collapsing back to a day. Nothing exotic, because nothing exotic is needed.

The two parts that are deliberate choices rather than algorithm are the ones the previous section names. Review items are generated from the lessons you finished, so there is no separate deck to maintain and no second app to open — they appear on the same daily plan as new lessons. And when you miss two weeks, nothing goes overdue: the schedule re-flows into the days you have left, which is the failure mode that ends most people's attempt at this.

You can also read the guide on sequencing a subject, which is the other half of the problem: spacing decides when material comes back, and prerequisite order decides whether it made sense the first time.