Spaced repetition for language learners, without the jargon

You learned el rompecabezas on Tuesday. On Friday it's gone, and a jigsaw puzzle is sitting on the table in front of you while your brain returns nothing.

Spaced repetition is the fix, and it's an old one. Ebbinghaus published the forgetting curve in 1885, and Wozniak's SM-2, the algorithm most spaced repetition software still inherits, is from 1987. It works. But almost everything written about it stops at the definition, which is the least useful part, because "review material at increasing intervals" tells you nothing about who decides the intervals or how.

So let's do the mechanism.

The two findings underneath it

Spaced repetition is 2 pieces of laboratory evidence wired to a calendar. Both are old and replicated. Boring, in the good way.

Spacing beats massing. Take a fixed amount of study time and spread it across days rather than piling it into one sitting, and you remember more. Cepeda and colleagues pulled together 317 experiments on this in 2006 and the effect held across the lot. The same hour, chopped up, buys more memory than the same hour in a block.

Retrieval beats review. Trying to produce an answer from your own head does more for memory than looking at the answer again. Karpicke and Roediger tested this on exactly the material you care about: Swahili-English word pairs. Their finding, in the paper's own words, is that repeated studying after learning "had no effect on delayed recall, but repeated testing produced a large positive effect." Rereading a word you've already got right once buys you nothing a week later. Being made to produce it buys you a lot.

The two results this whole method rests on

0

experiments on distributed practice reviewed by Cepeda and colleagues, 2006. Spreading study time out won

0

the year Ebbinghaus mapped the forgetting curve. Murre and Dros replicated its shape in 2015

Put those together and you get the whole design. Recall the word from memory rather than reading it, and do that on a schedule that spreads out over time. Everything else is bookkeeping about the dates. Both findings have their own posts here: the testing effect and the forgetting curve.

What an interval is

An interval is the gap in days between the last time you saw a card and the next time the software shows it. That's it. Whatever else a scheduler seems to be doing, the output is one number: how many days until this card comes back.

Under SM-2 at its starting ease, a new word runs 1 day, then 6, then 15, then 37, then 94. Each successful recall buys a longer gap. Miss the card and the gap collapses, the card comes back tomorrow, and it climbs again from lower down. Other schedulers pick different numbers, and the growing shape is the part they all share.

The reason the gaps grow rather than staying fixed is that memory decays fastest right after you learn something and then flattens out. Ebbinghaus mapped that shape in 1885, and Murre and Dros re-ran the experiment in 2015 with a proper sample and got the same curve back. A word you've successfully recalled 5 times over 4 months is decaying far more slowly than one you met yesterday, so it doesn't need checking as often.

An interval is also, deliberately, uncomfortable. A good scheduler shows you the card near the point where you're about to lose it, because a struggling recall builds more memory than a smooth one. That's desirable difficulty, and it's why well-scheduled reviews never feel easy. If your reviews feel comfortable, your intervals are too short.

What "adaptive" means, precisely

This is the word that gets dropped into a sentence and left there unexplained, so here's the definition with the vagueness taken out.

An algorithm is adaptive when your answers change the schedule. A card you keep fumbling develops shorter intervals than a card you nail every time, and both are on different tracks from your friend's copies of the same 2 cards.

The fastest way to see what that means is to look at a scheduler that isn't adaptive. Memrise runs a fixed interval ladder, published in its help centre: 4 hours, 12 hours, 24 hours, 6 days, 12 days, 48 days, 96 days, 6 months. Get it wrong and you drop back to 4 hours. Every user gets that ladder. Every item gets that ladder. Nothing you do moves a rung.

Clozemaster does the same thing with a mastery ladder: same day at 0%, next day at 25%, 10 days at 50%, 30 days at 75%, 180 days at 100%. Four correct answers in a row and the item is mastered; one wrong answer sends it back to 0%.

Both of those are real spaced repetition, and both work, and neither is adaptive. The schedule was decided in advance by a person and your performance only moves you along it. Which is a defensible design, honestly. It's simple and predictable. Nothing can go badly wrong with it.

An adaptive scheduler throws the ladder away and computes a date per card instead. Here's the whole comparison in one table.

One correction while we're here, because a lot of pages get it backwards. Duolingo does have published spaced repetition. Settles and Meeder presented half-life regression at ACL 2016, fitted to Duolingo's own practice logs, and open-sourced it. Whether that exact model still runs in production today isn't something I can verify from outside, so I won't claim it either way. The paper is real and it's good.

SM-2, in plain words

SM-2 is Piotr Wozniak's algorithm from 1987, written for SuperMemo, and it's still what Anki runs out of the box, which is why most people using spaced repetition today are on a 1987 algorithm without knowing it.

The idea fits in a sentence. Every card carries one number, called the ease factor, which starts around 2.5. The next interval is the last interval multiplied by the ease factor. Grade a card hard and its ease factor drops a bit, so its intervals grow more slowly from then on. Grade it easy and the ease factor climbs.

The first 2 intervals are hardcoded at 1 day and 6 days, and the multiplier only starts at the third review. So run the formula yourself and a card at ease 2.5 goes 1 day, 6 days, 15 days, 37 days, 94 days. Fail a card and the interval resets to the bottom while the ease factor takes a permanent knock.

A card you keep failing gets its ease knocked down each time, and because SM-2 floors ease at 1.3, a badly-behaved card ends up multiplying by 1.3 forever. 1 day, 6, 8, 10, 13. That's ease hell, in the words of the Anki users who kept hitting it: the card sinks to minimum ease and never climbs out.

What SM-2 doesn't have is any model of forgetting. It never asks how likely you are to remember a card today. It just multiplies. That's the honest limitation, and ease hell is what the limitation looks like from inside a deck.

Wozniak kept iterating. SuperMemo is on SM-19 as of 2024, having gone through SM-17 in 2015 and SM-18 in 2019. The versions after SM-2 aren't open, which is probably why the rest of the world stayed on the 1987 one for more than 3 decades.

FSRS, in plain words

FSRS, the Free Spaced Repetition Scheduler, was created by Jarrett Ye and built out by the Open Spaced Repetition community. It descends from MaiMemo's DHP model, and the underlying scheduling work was published at KDD 2022.

Instead of one number per card, it tracks 3. The definitions below are the project's own:

• Retrievability is "the probability that the person can successfully recall a particular piece of information at a given moment." It falls a little every day you don't see the card.
• Stability is "the time, in days, required for R to decrease from 100% to 90%." This is the number your reviews are trying to grow, from hours at the start to years eventually.
• Difficulty is "the inherent complexity of the information," per card, per learner. Stubborn cards grow their stability more slowly.

Two consequences follow from having a memory model rather than a multiplier.

First, FSRS schedules to a target you pick. Desired retention is adjustable from roughly 70% to 97%, and when FSRS is turned on in Anki the default is 90%. So the card arrives on the day your predicted recall has sagged to 9 in 10. Ask for 97% and you'll review far more often for a small gain; drop to 80% and your daily pile shrinks a lot while you forget more. SM-2 has no equivalent dial because it has nothing to predict with.

Second, FSRS-6 carries 21 parameters that get machine-fitted to your own review log. Two people can answer the same card the same way and get different next dates, because the model has learned different things about how each of them forgets. That's the sharpest version of "adaptive" available in shipping software today.

The usual claim you'll see is that FSRS needs 20 to 30% fewer reviews than SM-2 for the same retention. Worth flagging: that number comes from FSRS's own documentation rather than an independent trial. I use FSRS and I think the design is better. I should say that I build a product that runs FSRS, so read that preference with the appropriate suspicion. And I'd rather tell you where the figure came from than repeat it as though someone neutral measured it. The full mechanism is in the FSRS post.

Why your grades are the whole input

Every scheduler on that table, adaptive or not, sees your memory through one channel: the grade you tap after each card.

That makes self-grading the load-bearing part of the system, and it's where most decks quietly fall over. You see the answer, think "yeah, I basically had that," tap Good, and a word you half-knew gets scheduled like a word you own.

The research says your instinct here runs hot in a predictable direction. Koriat and Bjork's 2005 paper on illusions of competence describes the mechanism: judging how well you know something while the answer sits in front of you produces a number that's much too high. Everything looks knowable with the answer on the table.

The fix is one rule. Commit to your answer out loud, or type it, before you flip the card. A spoken guess can't be quietly upgraded after the fact. If the word only arrived after the flip, that's a fail, and marking it as one costs you 8 seconds tomorrow instead of 3 weeks of silent decay. The full rubric with the edge cases is in the grading post.

One more thing worth doing at the deck level: don't study one topic in a block. Mixing card types and topics within a session, which the literature calls interleaving, feels worse and tests better. Most spaced repetition apps shuffle by default, so you probably get this free.

What spaced repetition will not do

Here's the part that gets left out of pages selling something. I'd rather lose the sale than skip it.

Spaced repetition is a memory system. It builds reliable recall of items you have already met and understood. That's a narrow job, done extremely well. It's also nowhere near sufficient on its own.

A card teaches a word's core: form, meaning, sound. Everything around that core, it handles badly. Take collocation, which is the question of which words keep company with which. In Spanish you tomar una decisión. Say hacer una decisión, which is what English pushes you toward, and you've just marked yourself as a learner. No card taught anyone which verb goes there. Register is the same story, the line between what you say to your tutor and what goes in the group chat. So is sarcasm, where the sentence says one thing and is doing another. All of it lives in context and only arrives through volume.

There's a speed problem too. Real speech arrives fast, with no replay button. Catching words at that rate is a separate skill, trained only by hours of listening. A learner with a few thousand well-drilled words and no listening behind them freezes in a real conversation and then concludes they're bad at languages. They aren't. They built one component and skipped the rest.

So run input alongside the deck from day one. Reading, shows, podcasts, and talking to an actual person all count. A deck installs the units fast, and input is what teaches you what those units do. The argument for how to split your hours is laid out in comprehensible input vs. flashcards. If you only take one thing from this page, take that a deck without input is half a method.

Starting without overthinking it

The algorithm question matters less than the volume question, so here's the practical shape.

Add fewer new cards than you want to. Your daily review load is set by your new-card rate. Each card you add keeps coming back on a shrinking frequency for months, so today's 10 new cards are still generating reviews in November, on top of every batch you've added since. Ten a day sounds modest and turns into a review pile you did not sign up for. Five is plenty.

Make the card ask you to produce something. Put a gap in a sentence and make yourself fill it. Producing the word is the half that shows up when you're speaking.

Only make cards for words you've actually met. The best decks are harvested from lessons, shows, and books rather than downloaded as a 5,000-card list. A word you met in context has a hook already attached.

Do it daily, briefly. Five minutes a day beats an hour on Sunday, which is the spacing finding applied to your own calendar. Missing a day isn't a catastrophe, the cards just pile up slightly, and a scheduler is fine with being late.

You can run all of this on paper with a Leitner box: 5 physical boxes, a card moves up a box when you get it right and back to box 1 when you don't, and you review box 1 daily, box 2 every 3 days, and so on. That's a fixed ladder and it genuinely works. The thing software adds is knowing which 12 of your 600 cards are due today, which is exactly the calculation a human can't do in their head past a few weeks.

Some of you should go install Anki

If you like building your own decks, and you don't mind losing an evening to a deck options page, install Anki. Turn FSRS on at the bottom of that page and you're done.

It ships no language content of its own, which is the catch and also the point: you're buying an engine and supplying the Spanish yourself. The desktop app is free on Windows, macOS and Linux, AnkiDroid is free, and syncing through AnkiWeb is free. iOS is the only platform that costs anything, US$24.99 once, and that's a purchase rather than a subscription. For a certain kind of person, that's the correct answer to this entire page, and it's the one I'd give a friend who told me they enjoy making cards.

Questions people actually ask

---

Sources

Every load-bearing claim Verbamor makes is traced to its paper on the research page.