Spaced Repetition for Language Learners: A 2026 Guide

You've heard spaced repetition is the closest thing to a cheat code for memorizing vocabulary. You've maybe installed Anki, grabbed a deck, and then quit three weeks later buried under 800 reviews. The problem isn't the technique. It's that most learners are handed a tool without the model for when to use it, what to put in it, and when to let cards die. This guide walks through the research, the modern algorithms (including FSRS-6 and SuperMemo's new API), and a practical way to fit spaced repetition around native-content immersion instead of replacing it.

What spaced repetition actually is

Spaced repetition is the practice of reviewing information at expanding intervals, timed so each review happens just before you would have forgotten. The core intuition comes from Hermann Ebbinghaus's 1885 monograph Über das Gedächtnis, where he tested himself on roughly 2,300 nonsense syllables and plotted the now-famous forgetting curve. Memory decays quickly at first, then flattens. If you restudy right before forgetting, the curve resets and decays more slowly each time.

In 2015, Jaap Murre and Joeri Dros at the University of Amsterdam replicated Ebbinghaus's experiment and confirmed the curve holds, with a small upward bump around the 24-hour mark that they attributed to sleep-dependent memory consolidation. This matters for a practical reason: if you cram the night before a test, you're fighting the curve. If you review tomorrow, the day after, then four days later, you're surfing it.

The effect isn't subtle. Herbert Spitzer's 1939 classroom study, published in the Journal of Educational Psychology, tested 3,605 sixth-graders in Iowa and showed massive retention gains from spaced review versus massed study. Nicholas Cepeda and colleagues' 2006 meta-analysis in Psychological Bulletin pulled together 839 effect-size contrasts from 317 experiments and found distributed practice reliably outperformed massed practice across subjects, ages, and formats.

The rule of thumb from Cepeda et al.'s 2008 follow-up (over 1,350 participants): the optimal gap between reviews is roughly 10 to 20 percent of how long you want to remember the item. Want to know a word a year from now? Space your reviews by about a month or two at steady state.

Why it works: testing, not just spacing

A subtle point most beginners miss is that spaced repetition is really two ideas fused together: the spacing effect, and the testing effect. Henry Roediger and Jeffrey Karpicke's 2006 study at Washington University in St. Louis found that students who studied a passage once and then took three recall tests retained about 60 percent of the material a week later. Students who studied the passage four times retained about 40 percent. Retrieval beats rereading.

Karpicke and Roediger followed up in Science in 2008 with a foreign vocabulary study showing that once a learner could recall a word correctly, additional testing contributed far more to long-term retention than additional study. The practical translation: flipping through a wordlist is close to useless. Forcing your brain to produce the word from a prompt, even when it hurts, is what builds the memory.

There's suggestive biology behind this too. Esteban Kramár and colleagues at UC Irvine showed in 2012 that theta-burst stimulation to rat hippocampal synapses produced no extra potentiation when repeated within minutes, but produced further strengthening when the bursts were spaced about an hour apart. Heather Sisti, Anthony Glass, and Tracey Shors at Rutgers (2007) found that spaced training in rats rescued newborn neurons in the dentate gyrus from programmed cell death in a way massed training didn't. You can reasonably take from this that spacing isn't a productivity hack. It's closer to how the tissue wants to work.

The algorithms: SM-2, FSRS-6, SM-20

Piotr Woźniak started his paper-based SM-0 system on July 31, 1985, and wrote Algorithm SM-2 in Turbo Pascal 3.0 in late 1987. He placed SM-2 in the public domain as part of his 1990 master's thesis, which is why SM-2 became the default engine of roughly every open flashcard app for the next three decades, including older versions of Anki.

SM-2 is simple and conservative. You grade a card 0 to 5, it adjusts an ease factor, and it picks a next interval. It works, but it's lossy. It assumes a fixed memory model that doesn't fit every learner or every card. It also has no concept of card difficulty beyond the ease factor, which means a card that keeps failing just gets its interval shrunk without the algorithm actually learning that the card itself is hard.

FSRS, the Free Spaced Repetition Scheduler, is the modern replacement. FSRS-6 shipped in late 2025, trained on roughly 700 million reviews contributed by about 20,000 volunteer Anki users. It models three variables per card: stability (how long the memory lasts), difficulty (how hard the card is for you specifically), and retrievability (your probability of recall right now). Benchmarks indicate it reduces the number of reviews required by 20 to 30 percent compared to SM-2 at a 90 percent retention target. For a learner doing 200 reviews a day, that's 40 to 60 reviews a day you get back, permanently. From Anki 23.12 onward, FSRS is available in the scheduler, and newer installs default to it for fresh profiles (existing profiles stay on SM-2 unless you flip the switch). If you've been on Anki for years and never touched the setting, you're almost certainly leaving efficiency on the table.

On the commercial side, SuperMemo launched a public API on March 31, 2026, exposing the SM-20 algorithm to developers. The early-access tier allows up to 100 repetitions per day and a one-time import of up to 10,000 historical repetitions. SM-20 incorporates what Woźniak calls the two-component model of memory, tracking retrievability and stability as separate quantities much like FSRS does, though with different training data and a different smoothing approach. That matters less for the individual learner than for what it signals: the next few years of language apps will have much smarter schedulers baked in, not just hand-tuned intervals.

Which algorithm should you actually use

If you're on Anki, turn on FSRS. It's free, it ships with the app, and the efficiency gain is real. Set your desired retention between 0.85 and 0.90 and optimize the parameters after you have a few hundred reviews in your history. If you're on a modern app like Migaku, the scheduler is already FSRS-based and you don't need to think about it. If you're on an older tool still running SM-2, the honest move is to migrate. The algorithm gap is large enough that you're paying for it in time every single day.

What to actually put in your deck

Algorithms schedule cards. They don't pick them. This is where most learners silently fail. A deck full of 10,000 cards scraped from a frequency list is a chore. A deck full of 2,000 cards from sentences you actually read or heard is a memory.

A workable policy:

For Japanese specifically, mining from shows and manga tends to outperform premade decks past the first thousand words. The same logic applies to most languages (see what actually works for learning Japanese for a worked example).

Fitting SRS into an immersion routine

Spaced repetition is a support system. It's not the main event. The main event is hours spent reading, watching, and listening to things you find interesting. SRS exists so that words you met in content don't slip away before you meet them again.

A realistic daily loop for an intermediate learner:

This pattern is roughly what strong self-learners converge on after a year or two of experimentation. It's also the logic behind how Migaku works: reading and watching come first, the flashcards are generated from what you actually encountered, and an FSRS-style scheduler handles the timing.

If you're earlier in the journey or picking a language, the broader context in how to actually learn a language and what makes a language easy to learn is worth reading before you commit to 500 hours of anything.

A worked example: one week of mining

To make this concrete, here's what a week of disciplined mining looks like for a Spanish learner at roughly B1 level watching La Casa de Papel.

Monday, 45 minutes of viewing. Eight unknown words flagged: atraco, rehén, imprenta, sucursal, pasmado, cuartel, desvío, trastero. Five go into the deck as sentence cards. Pasmado and trastero feel too rare for episode one, so they get skipped. Imprenta already appeared in a podcast last week, so it jumps the queue.

Tuesday, 30 minutes of podcast listening plus a graded reader chapter. Six new words, three cards created. By Friday, atraco has appeared in three separate contexts, which is the point at which words genuinely stick without much conscious effort. The Monday card for atraco shows up for review on Wednesday, again the next Sunday, and then not for two weeks. Each time it comes up, it takes about three seconds because the show's opening scene loads alongside the word.

By Sunday, the week has produced 18 cards from roughly 6 hours of content. That's a sustainable pace. Over a year, it's somewhere around 900 to 1,000 new words, all of them anchored to real memories. Compare that to grinding a 10,000-word Anki deck cold, where retention drops fast and motivation drops faster.

Cultural and linguistic context matters

One thing the algorithms can't model is how differently vocabulary loads depend on the language. Japanese kanji have reading ambiguities, so a single card often needs to teach three things at once: the kanji form, the reading, and the meaning. A word like 生 can be read sei, shou, nama, i(kiru), u(mareru), or ha(eru) depending on context. Building a card that tries to drill all six readings at once will fail. The workable move is to mine the specific compound you encountered, like 生物 or 生まれる, and trust that other readings will show up in other sentences later.

Spanish verbs have conjugation tables that a single dictionary form card won't capture, so many learners mine whole conjugated phrases. A card for hubieran llegado teaches tense, mood, person, and number all at once. A card for llegar alone teaches none of them. German separable verbs break apart in sentences, so a card showing only aufstehen misses the fact that ich stehe um sieben auf is where the grammar actually lives.

This is why rigid one-size-fits-all deck policies tend to fail. The principle is constant (retrieve in context, space the reviews, keep the load modest), but the card format should match the language's actual difficulty surface. Learners of tonal languages like Mandarin or Vietnamese often add audio to every card because a silent reading of the tone is almost useless; a learner who can read 买 but not distinguish mǎi from mài in speech has built only half a memory. Learners of Arabic often mine whole sentences with full vowelization on the back, even though unvowelized text is what they'll see in the wild, because the vowel marks are what disambiguate otherwise identical consonant skeletons.

There's also a cultural register problem algorithms cannot see. Japanese has sharply different speech levels (plain, polite, humble, honorific), and a word learned in a reality show may be exactly wrong in a business email. Korean has the same issue magnified. French tu versus vous shifts whole verb paradigms. Good mining pays attention to who said the line in the source material, not just what was said, and cards benefit from a quick note about register when the word is marked for it.

Common mistakes to stop making

A short list of things that quietly kill SRS routines:

How long until spaced repetition actually pays off

The honest timeline for a learner doing 10 to 15 new cards a day with consistent immersion looks something like this. In the first month you feel overwhelmed because everything is new and nothing has intervals yet. In the second month the intervals stretch out, the daily load stabilizes around 60 to 100 reviews, and you start recognizing your mined words in new content. By month six you have roughly 1,500 to 2,500 mature cards, the daily review load is still around 80 to 120 cards, and content you found impossible in month one is merely hard. By year two you stop thinking about the deck as a project and start thinking about it as a maintenance habit, like brushing teeth.

The tempting mistake at every stage is to accelerate. Double the new cards for a week, and a month later the review load is unmanageable and you quit. The learners who last are the ones who treat the daily number as a hard cap, not a floor. FSRS rewards consistency far more than intensity.

Frequently asked questions

How many cards per day is realistic for a working adult?

For someone with a full-time job, 10 new cards a day is sustainable long term. That produces a steady-state review load of roughly 80 to 120 cards a day after a few months on FSRS, which fits in a 20-minute slot. Going higher works for a few weeks, then collapses. The learners who still have their deck active after two years almost always started small.

Should I use a premade deck or build my own?

For the first 1,000 to 2,000 most-frequent words, a premade deck is fine and saves time. After that, switch to mining. Premade decks past the core vocabulary tend to be full of words you'll never encounter, and cards without personal context are the hardest to remember.

What do I do with a word I can't remember no matter how many times it comes up?

Suspend the card and stop trying. The word is signaling that you haven't met it enough times in real content for the memory to have any hooks. When you meet it again in a show or book, you can either remake the card with better context or just notice it and move on. Memories form when there's something to attach them to.

Is it okay to skip a day?

One day, yes. FSRS and SM-2 both handle a one-day gap gracefully. A week is where the backlog starts to hurt. If you know you'll be away, set new cards to zero a few days in advance so the queue drains, then resume fresh when you're back.

Does spaced repetition work for grammar, or only vocabulary?

It works for anything you can turn into a cue-and-recall pair. Sentence cards with a grammar point highlighted are a strong format for patterns like Japanese particles, Spanish subjunctive triggers, or German case endings. The trick is that the card should test production or recognition of the pattern in context, not a rule in the abstract. A card that asks "when do you use the subjunctive?" is close to useless. A card that shows espero que ___ bien and expects estés is useful.

Can I use spaced repetition for listening or pronunciation, not just reading?

Yes, and most serious learners should. An audio card plays a sentence and asks you to either transcribe it or state the meaning before seeing the text. This trains the ear in a way reading cards never will. For tonal languages it's close to mandatory. For any language it closes the gap between silent recognition and real-time comprehension.

The learners who make real progress aren't the ones with the biggest decks or the fanciest setups. They're the ones who show up, keep reviews modest, and spend most of their language time inside content they'd consume anyway.

If you want spaced repetition to quietly handle itself while you focus on the shows, books, and articles you actually care about, Migaku is built around exactly this loop: hover to understand, one click to mine, and an FSRS scheduler keeping the reviews light.