How spaced repetition actually works

The shape of forgetting

In 1885 a German psychologist named Hermann Ebbinghaus memorised lists of nonsense syllables, retested himself at intervals, and plotted what he could still recall. The curve he drew has been replicated thousands of times since. It is steep at the start — most of what you have just learned drains away within a day — and then flattens. The second time you encounter the same item, after the first wave of forgetting has happened, the curve is shallower. The third time, shallower still. By the fourth or fifth review the slope is so gentle that you can leave a fact alone for months and still find it intact.

Spaced repetition is the trivial-sounding consequence: review the item right before you would otherwise forget it. Each well-timed review pushes the next safe interval out further. Eight reviews, scheduled correctly, will keep a fact retrievable for years. Eight reviews crammed into a single afternoon will not.

This is not a hack and it is not new. It is the only memory technique with a hundred and forty years of evidence behind it. What has changed in the last fifty is the willingness of computers to track which item you saw when, predict when each one is about to fall off the edge, and put exactly those items in front of you today.

Why “just review every day” is wrong

The intuitive alternative to spaced repetition is to review your whole vocabulary list daily. People do this with paper flashcards and Quizlet decks all the time. It feels thorough. It is enormously wasteful.

Half the items on a daily list are ones you already know cold — reviewing them adds nothing and consumes the time you could have spent on the half you don't. The other half are items you forgot weeks ago, where one repetition does not patch the memory; you need a few well-spaced ones, which the daily-list approach cannot give you because it scatters every item across the same uniform interval.

Spaced repetition fixes both halves at once. Items you got right yesterday won't reappear today. Items you got wrong yesterday will. Items you have aced four reviews in a row drop into a months-long quiet zone. The deck is sorted by who-needs-help-now, not by alphabetical order.

The two algorithms HanziFluency runs

Two scheduling algorithms ship with the site: SM-2 and FSRS-5. Both are based on the same forgetting curve. They differ in how aggressively they try to learn from your specific review history.

SM-2 — the workhorse

SM-2 is the algorithm that powered SuperMemo in the late 1980s and that Anki defaulted to until recently. It is a small set of rules: every card has an interval and an “ease factor”. When you get a card right, the next interval is the previous one multiplied by the ease factor (typically around 2.5). When you get one wrong, the interval resets and the ease factor drops a little. That is genuinely the whole thing.

SM-2's strengths are robustness and intelligibility. It works fine from your very first review, with no warm-up data, and you can predict its behaviour in your head. Its weakness is that the ease factor is a blunt instrument: it cannot tell the difference between a card you find genuinely hard and one you happened to fail when tired, and it cannot adjust the multiplier per-card based on which intervals have actually been working for you.

FSRS-5 — the data-driven option

FSRS (Free Spaced Repetition Scheduler) replaces the ease-factor heuristic with an explicit memory model: every card has a difficulty, a stability, and a retrievability, and the algorithm fits seventeen parameters to your past reviews so the next interval matches your actual forgetting rate. Modern FSRS implementations beat SM-2 by ten to thirty per cent on retention-per-review in published comparisons.

The catch is that FSRS only starts to outperform SM-2 once it has enough of your review history to fit. The HanziFluency implementation runs with sensible defaults out of the box and refines them in the background as you accumulate reviews; for the first week or two on a new account, it will behave roughly the same as SM-2.

Practical recommendation: leave SM-2 on for your first month of study, then switch to FSRS-5 once your review log has some shape to it. The toggle is in Settings and your historical data is preserved on either side of the switch.

The retention rate dial

Both algorithms exist to hit a target retention rate. The default is around 90% — the algorithm tries to schedule each review so you have a 90% chance of getting it right. Lower the target and the intervals stretch (fewer reviews, more failures). Raise it and the intervals shrink (more reviews, fewer failures).

90% is the right number for most learners. It produces enough successful retrievals to feel confident, while leaving enough difficulty in the deck that you are genuinely strengthening memories rather than just rehearsing things you already know. Drop it to 80% if you are happy with more failures and want fewer reviews per day; raise it to 95% if you are about to take an exam and need the last few percent of certainty.

Where spaced repetition quietly fails

SRS is excellent at one thing: keeping a known item retrievable. It is mediocre or useless at three other things, and learners who lean on it for those things end up frustrated.

• Acquisition. SRS keeps memories alive; it does not create them. The first encounter with a character — seeing it written, hearing the pronunciation, attaching it to a meaning — is a separate act, and SRS is just the maintenance system that runs afterwards. Trying to learn a brand-new character solely from a flashcard prompt is the slow, unpleasant way.
• Production. Recognising 谢 on a card is a different skill from producing xièxie in conversation when somebody hands you a coffee. SRS drills the recognition; the production has to come from speaking practice. The HanziFluency writing and stroke-recall drills push partway towards production for handwriting; for spoken production you need a tutor or a language partner.
• Listening at speed. SRS prompts are isolated, slow, and generously timed. Real listening is none of those things. Use the dictation and listening drills as a bridge, but real comprehension comes from podcasts, TV, and conversations.

In short: SRS is the gym. It builds and maintains a strength. The match itself happens elsewhere.

The operator decisions that matter

Most failure modes in SRS are not algorithmic — they are behavioural. A few rules of thumb that consistently distinguish learners who stick with it from learners who don't:

1. Don't binge. Sitting down to clear a 200-card backlog after skipping a week is the worst possible use of an SRS. The algorithm assumes you are spreading reviews across days; collapsing them all into one session ruins the spacing for everything you fail. Better to do thirty cards a day for a week than two hundred in one sitting.
2. Be honest with the grading. The temptation when you half-remember a card is to mark it as “good” because you nearly got there. Resist it. The algorithm needs accurate signal — calling a near-miss a hit shortens the next interval not at all and you will see the same near-miss again in two weeks instead of two days, which is when you will actually forget it.
3. Cap new cards. Every new card is a future review obligation — roughly seven to ten reviews over its lifetime, at a guess. Adding fifty new cards a day during an enthusiastic week produces a five-hundred-card daily review load three weeks later. The daily new-card cap is the single most underused setting in any SRS app. Set it lower than you think.
4. Lapse honestly, then move on. When you forget a card you “should have known”, don't spiral. Mark it wrong, look at the answer, move to the next one. The algorithm will surface it again tomorrow. The worst thing you can do is sit there ruminating, because that erodes the motivation to come back tomorrow at all.

The streak trap

Most SRS apps display a daily streak counter. HanziFluency does too, on the home page. Streaks are good motivational tools when they nudge you to do today's twenty cards instead of skipping. They are bad ones when they bully you into doing fifteen minutes of reviews at one in the morning, in the dark, on your phone, half asleep.

The data in those reviews is garbage. The algorithm interprets your sleep-deprived misses as evidence that the cards are harder than they are, then schedules them more aggressively for the next month. Better to break the streak. The cards do not care.

When you are done

You will know spaced repetition is doing its job when reviewing a Mandarin character feels less like work and more like a quick check-in with an old acquaintance. The daily review queue stays manageable — typically thirty to sixty cards once your deck matures — and the cards that surface are mostly the ones that genuinely benefit from another look. New characters move in on a steady drip rather than a flood.

At that point the algorithm has receded into the background, where it belongs. You are no longer thinking about intervals or ease factors. You are just learning, with a small, well-tuned engine quietly making sure none of it leaks out.

Next: reading vs. writing Chinese characters, or jump straight into your due reviews.