Spaced Repetition in 2026: What Actually Works for Language Learners

You've probably heard spaced repetition described as the closest thing language learning has to a cheat code. That framing sets the wrong expectation. Spaced repetition is a scheduling technique, and on its own it does not teach you a language. What it does, very well, is keep vocabulary and grammar patterns you've already encountered from falling out of your head. This article walks through what the research actually says, how modern algorithms like FSRS-6 and SM-20 differ from the 1980s-era SM-2, and how to fold all of it into an immersion-based routine without drowning in reviews.

What Spaced Repetition Actually Is

The core finding is almost a century old. In 1939, Herbert F. Spitzer tested 3,605 sixth-graders across Iowa on their retention of short factual articles about peanuts and bamboo, and published the results in the Journal of Educational Psychology. Students who reviewed the material at spaced intervals remembered far more than students who crammed. The effect has been replicated relentlessly since. Cepeda and colleagues' 2006 meta-analysis in Psychological Bulletin pulled together 839 effect-size contrasts from 317 experiments across 184 articles, and found the same thing every time: distributing practice beats massing it.

The mechanism behind this is the forgetting curve, first sketched by Hermann Ebbinghaus in the 1880s and replicated in 2015 by Murre and Dros at the University of Amsterdam (one subject, 70 hours of testing, curve closely matching the original, plus a small bump at 24 hours that the authors attribute to sleep consolidation). Memories decay predictably. Each time you successfully retrieve a memory just before it would have faded, the decay slope gets shallower. Spaced repetition software (SRS) is simply bookkeeping: it tracks when each fact is about to be forgotten and surfaces it at that moment.

One related idea worth separating out is the testing effect. Roediger and Karpicke's 2006 study at Washington University in St. Louis found that after one week, students who practiced active retrieval recalled roughly 60% of material, versus roughly 40% for students who only re-read it. This is why SRS uses flashcards rather than passive re-reading. The retrieval itself is part of the medicine.

The Algorithms: SM-2, FSRS-6, and SM-20

Piotr Woźniak wrote the first computerized spaced repetition algorithm, SM-2, in Turbo Pascal 3.0 over sixteen evenings in late 1987. He placed the full description in the public domain as an appendix to his 1990 master's thesis, where he reported memorizing 10,255 English vocabulary items in the first year of using it, at 41 minutes per day, with 92% retention (excluding items in their first three weeks). That algorithm, essentially unchanged, is still what most SRS apps fall back to.

SM-2 has well-known problems. In Anki's implementation, every time you press "Again" on a card, its ease factor drops by 20%. Cards that you fail a few times in a row get stuck at ease factors around 130%, where their intervals barely grow. The Anki community calls this "ease hell." You end up seeing the same stubborn cards every other day forever.

FSRS (Free Spaced Repetition Scheduler) addresses this directly. FSRS-6 shipped in late 2025 and was trained on roughly 700 million reviews contributed by about 20,000 volunteer Anki users. At a matched 90% retention target, it reduces the number of reviews required by 20 to 30% compared to SM-2. As of Anki 23.12 and later, FSRS is available as the scheduler, and it's the default for new profiles on recent installs. If you're still on SM-2 out of inertia, switching is one of the higher-leverage moves you can make.

SuperMemo has continued its own line of development. SM-20, announced in 2026, is the first SuperMemo version where all scheduling parameters are computed by machine learning rather than hand-tuned heuristics. The SuperMemo API launched in early access on March 31, 2026, exposing SM-20 to developers with free usage limits of 100 repetitions per day and a one-time import of up to 10,000 historical repetitions. Expect to see SM-20 showing up in third-party tools over the next year.

The practical takeaway for a language learner: you do not need to care deeply about which algorithm your app uses, as long as it's FSRS or something newer. The differences between FSRS and SM-20 are small compared to the difference between using any modern scheduler and not doing spaced review at all.

How FSRS Actually Differs Under the Hood

It helps to understand, at a high level, why FSRS behaves better than SM-2 in practice. SM-2 has two state variables per card: an interval and an ease factor. It updates them with a fixed formula regardless of your own review history or the specific card's difficulty. FSRS uses a three-component memory model borrowed from the cognitive science literature: stability (how slowly the memory decays), difficulty (how hard the item is for you specifically), and retrievability (the probability you'd recall it right now). These are updated with parameters fit to your own review log once you have around 1,000 reviews on the books.

The concrete consequences:

For most learners this translates into roughly a quarter fewer daily reviews for the same recall, which compounds over the years you'll actually be using the system.

Optimal Intervals (and Why the Defaults Are Usually Fine)

The most useful concrete finding on intervals comes from Cepeda et al.'s 2008 follow-up, which tested over 1,350 participants and found that the optimal gap between study sessions scales to roughly 10 to 20% of the desired retention interval. If you want to remember a word for a year, the gap between successful reviews should eventually stretch to about 5 to 10 weeks. If you want to remember it for a week, gaps of a day or so are appropriate early on.

Modern schedulers handle this math for you. What you control is the retention target. FSRS lets you pick it: 90% is the default, 95% means shorter intervals and more daily reviews, 85% means longer intervals and more forgetting. For a language learner reviewing 20 to 40 new cards a day, 90% is almost always the right setting. Pushing to 95% roughly doubles your review load for marginal retention gain. Dropping to 85% saves time but you'll fail more cards, which feels worse and can sap motivation.

A few patterns that show up in practice:

Where SRS Fits in an Immersion Routine

Here's where the method gets misused. Learners hear about spaced repetition, download Anki, grind through a pre-made 10,000-word deck, and wonder why they still can't understand a podcast. The reason is that an SRS card of a word you've never heard in context is close to useless. You're memorizing a gloss, not a word.

The version that actually works looks more like this. You watch an episode of Terrace House or read a chapter of Kiki's Delivery Service. You hit a word you don't know, look it up, and if it seems useful you make a card from the exact sentence you saw it in. The card has the sentence on the front, the target word bolded, and the audio if available. The back has the definition and a screenshot or image. When the card comes up in review two days later, you're not recalling an abstract dictionary entry. You're recalling the scene.

This is the sentence-mining approach, and it's been the standard among serious immersion learners for over a decade. For a broader treatment of how to structure a full routine around this, How to Actually Learn a Language walks through the daily schedule. For Japanese specifically, which has its own quirks around kanji and pitch accent, How to Learn Japanese Practically covers what to mine and when.

A reasonable daily split for an intermediate learner:

Notice the ratio. SRS is maybe 20% of your time. The content is the meal. Flashcards are the vitamins.

Beyond Vocabulary: Spacing Grammar, Listening, and Output

Spaced repetition isn't only for words. Grammar patterns respond to the same treatment when you card them as example sentences. A card for the Japanese ～てしまう pattern might look like:

Front: the sentence with audio. Back: the English gloss and a one-line note on what ～てしまう conveys (completion, regret, unintended result). You're not memorizing a rule. You're memorizing a concrete instance, and the rule falls out of seeing enough instances.

Listening deserves its own spaced approach. One trick: when a card's audio gives you trouble, don't just grade it and move on. Shadow the sentence three or four times before clicking through. This layers pronunciation practice onto your review queue without adding a separate study block. For a deeper treatment of the technique, Shadowing as a Learning Technique gets into the mechanics.

Output (speaking and writing) is the one area where SRS helps least directly. You can card production prompts ("how do you say X?"), but the real gains come from conversation practice. Use SRS to keep vocabulary ready for retrieval, then use that vocabulary in actual exchanges.

Card Design: Small Choices That Compound

The difference between a deck that feels maintainable at year three and a deck you abandon at month four is almost entirely about card design. A few principles that hold up across languages:

A small test: pick ten of your oldest cards and ask whether you remember where each sentence came from. If you can picture the scene or the page, the card is doing its job. If the sentence feels context-free, it probably is, and it will leech.

Cultural and Language-Specific Wrinkles

SRS advice tends to be written as if all languages were the same. They aren't, and a few common situations deserve calling out:

Worked Example: A Week of Mining from One Show

To make this concrete, here's what a week of realistic SRS usage looks like for an upper-beginner Spanish learner watching La Casa de Papel.

Monday: 45 minutes of one episode. Seven lookups, four of which become cards. Two minutes per card to set up front, back, audio, and screenshot. Total SRS time: 8 minutes adding, 12 minutes reviewing older cards. 20 minutes of SRS, 45 minutes of content.

Wednesday: Another episode. Six new cards. The Monday cards reappear for their first review. Two of them are graded "Good," one "Hard," one "Again." The one you failed had two new words in it, which is the lesson: next time, pick a simpler sentence.

Friday: Lighter day. No new episode, just reviews. 25 minutes total. Monday's cards are now at 6-day intervals, Wednesday's at 3 days.

Sunday: One more episode, five new cards. Review queue is around 35 cards. Everything fits in 30 minutes.

Over the week: four episodes watched, 22 new cards added, roughly 140 reviews completed. Compare this to a learner doing 50 new pre-made cards a day, who ends the week with 350 cards and zero episodes watched. The mined-card learner has worse raw numbers and better actual comprehension.

Common Failure Modes

A few ways this goes wrong, ranked by how often we see them:

The underlying point: SRS is scheduling. It schedules whatever you put into it. If you put in cards mined from content you love, it keeps that content accessible. If you put in a context-free word list, it keeps a context-free word list half-memorized.

Frequently Asked Questions

How many cards a day should I add as a beginner?

Five to ten. The review load compounds: ten new cards today means roughly forty reviews due in a week and seventy in a month once cards start reappearing at mixed intervals. Beginners routinely set thirty or fifty new cards a day, feel fine for two weeks, and then quit when the review queue hits 300. Start small, feel the load settle, and raise the number only when reviews fit comfortably into your day.

Should I use a pre-made deck or mine my own cards?

A small pre-made frequency deck (the first 1,000 to 1,500 words) can be a reasonable shortcut to get you to a point where immersion is productive. After that, mined cards outperform pre-made ones because they're linked to scenes and sentences you actually remember. The hybrid pattern that works: pre-made deck for the first couple of months, then taper it off as your own mined deck takes over.

What do I do when my review queue gets out of control?

Stop adding new cards entirely until the queue is clear, which usually takes one to two weeks. Do not try to "catch up" by doing 400 reviews in one sitting; you'll grade sloppily and corrupt the scheduler's data about your actual recall. Cap reviews at your normal daily number plus 30%, and accept that clearing takes time. If the queue is over a thousand and the cards no longer feel like yours, it's often faster to reset mature-card intervals and treat them as relearning than to grind through.

Does spaced repetition work for adults as well as children?

Yes. The Cepeda meta-analysis covered participants from age 5 to over 70, and the spacing effect shows up at every age tested. What changes with age is raw encoding speed, not the benefit of spacing. Adults often do better with SRS than children because they're more consistent with daily review, which is the single biggest factor.

Can I use SRS for speaking practice?

Indirectly. SRS keeps vocabulary and set phrases available for retrieval, which is a precondition for fluent speech, but it doesn't build the motor and timing skills of actual conversation. A workable pattern: use SRS to maintain vocabulary, and spend separate time on output through tutoring, language exchange, or writing practice. Production cards ("how do you say X in your target language?") can bridge the gap but shouldn't replace real conversation.

How long until I notice results?

For vocabulary recognition in context, two to three months of consistent daily review (twenty minutes, mined cards, daily content). You'll notice it first when a word you carded three weeks ago shows up in a new episode and you understand it without stopping. For broader comprehension gains, six months is a more honest timeline. Anyone promising faster is selling something.

Is Anki still the best tool, or has something replaced it?

Anki is still the bedrock because it runs FSRS, is free, and has a massive ecosystem. The weakness is card creation: building good sentence cards by hand is slow enough that most learners give up. Tools that integrate lookup, audio capture, and card creation into the content-consumption step solve this friction, which is why the modern workflow looks more like "watch show, one-click card" than "open Anki, type fields."

Migaku was built around the sentence-mining loop described above: you watch or read native content, hover to look up anything you don't know, and turn useful sentences into SRS cards with one click, audio and screenshot included. If you want to see how the immersion side and the review side fit together in one workflow, take a look at how Migaku works.