The best spaced repetition apps, and how their schedulers differ

If you search this query you'll find roundups that hand every app a retention percentage. The one ranking above me, laxuai.com's 2026 roundup, studied a 50-card pharmacology deck for 14 days and reports 89% for Anki against 87% for Laxu AI, the app it sells. No raw data, no named tester, no sample size, no screenshots.

One person reviewing one deck for 2 weeks cannot tell 87% from 89%. That gap is smaller than the noise from a single bad night's sleep, and the whole ranking rests on it.

The tell is in the same page's algorithm column, which calls Anki "SM-2 (customizable)" and RemNote an "SM-2 variant." Both of those apps have shipped FSRS for years. Whoever ran the 14 days didn't open the deck options screen.

What I did not do

I did not run a head-to-head retention test, and I'm not going to pretend I did. Doing it properly needs randomized assignment, matched material, hundreds of learners per arm and a delayed test. That's a research project with an ethics board, not a blog post with a spreadsheet.

What I did do is read the documentation and the source, then run FSRS-6 myself on its published default parameters, so the worked example below has real intervals in it instead of adjectives. Every scheduler claim comes from the app's own manual, help centre or code, with the source named, and I checked all of it on 11 August 2026.

I've named the page I'm rebutting and linked it, because quoting a rival's figures while hiding whose figures they are is the same unfalsifiability I'm objecting to. Go read it and decide whether I've characterised it fairly.

I have a stake in this. Verbamor is mine, it's on this list, and I've put it last on purpose.

The four kinds of scheduler

Strip away the branding and there are 4 things an app can be doing when it decides your next review:

• A fitted memory model. FSRS, and SuperMemo's later algorithms. Each card carries its own state, and the model's parameters are fitted to your review history. Adaptive per card and per person.
• A fixed formula with one adjustable number. SM-2, from 1987. Every card has an ease factor that drifts up when you pass and down when you fail. Your grades move that one number, and the formula around it stays the same for everybody.
• A fixed ladder. Memrise and Clozemaster. The intervals are hard-coded. Pass and you climb a rung, fail and you drop to the bottom. Nothing about you or the item changes the numbers.
• Nothing scheduled at all. Quizlet's Progress. It tracks how often you were right and never computes a next-review date.

All 4 are marketed with the same phrase. The reason it matters is that the first 2 aim your reviews at the moment a specific memory is fading, and the last 2 can't, because they don't model the memory.

One hard word, two schedulers

Here's the difference as it happens to an actual word. Take Spanish el sacacorchos, corkscrew. You've never held onto it.

Memrise's rungs are documented: 4 hours, 12 hours, 24 hours, 6 days, 12 days, 48 days, 96 days, 6 months, with a wrong answer resetting to 4 hours. So say you pass the first four reviews, fail a week in, then pass slowly at day 14.

Walk the rungs and the dates fall out. Review 1 at 4 hours, review 2 at 12 hours, review 3 at 24 hours, which is day 1. Review 4 is 6 days after that, so day 7, and that's where sacacorchos goes. The reset drops you to 4 hours and you climb 4h, 12h, 24h, 6d again, which puts the next test at roughly day 14.

Then I ran that same sequence of answers, on those same dates, through FSRS-6. Both columns below are arithmetic you can redo.

One hard word, days 0 to 14. Ladder rungs are Memrise's documented sequence. The model column is FSRS-6, default parameters, desired retention 0.9, before any personal fitting.

What happened	Fixed ladder	FSRS-6 on defaults
Day 0, learned it	Next review in 4h.	Difficulty 4.88, stability 3.26 days. Left to itself, it would ask again on day 3.
Reviews at 4h, 12h and 24h, all passed	4h, 12h, 24h, then the 6-day rung.	Stability creeps 3.26 → 3.83 → 4.94 → 6.55 days. Three reviews inside a day buy almost nothing, and the model prices them that way.
Day 7, failed	Back to 4h. Climb the same rungs again.	Difficulty 4.84 → 7.21, stability 6.55 → 1.69 days. Due about 2 days later. It records the failure instead of only reacting to it.
Day 14, passed but slowly	Next rung is 12 days, then 48.	Difficulty 7.82, stability 3.35 days, so the next review is about 3 days out. A high-difficulty card grows its interval slowly.
The word you never miss, same dates	Identical rungs: 4h, 12h, 24h, 6d, 12d, 48d.	Stability 39.6 days by day 14, so the next review is about 40 days out. Same answers in, 12x the interval out.

The failure cost the ladder the whole climb over again, and after a pass at day 14 it still hands you a 4x jump from 12 days to 48. FSRS charged the same failure to the card and kept it close, at roughly 3 days, while pushing the word you never miss out past a month.

That last row is the one I'd sit with. The ladder gives sacacorchos and gato the same 8 rungs forever. Count the rungs below 12 days and there are 4 of them, so every word you already own still costs you 5 reviews before it clears the 12-day mark.

For the model's behaviour over a longer stretch, with your grades driving the dates instead of a ladder's, the FSRS post follows one card through 6 weeks.

SM-2 and FSRS, mechanically

This is the distinction most pages on this query duck, so here it is in full.

SM-2 is Piotr Wozniak's algorithm from 1987, written for SuperMemo and inherited by nearly everything since. It applies a fixed formula. Each card holds one adjustable number, the ease factor, which rises when you grade well and falls when you don't, and the next interval is the previous one multiplied by that ease. Your review history changes the ease factor. It never changes the formula.

FSRS, the Free Spaced Repetition Scheduler, models each card with 3 quantities, the DSR model. The Open Spaced Repetition wiki defines them precisely:

• Retrievability is "the probability that the person can successfully recall a particular piece of information at a given moment."
• Stability is "the time, in days, required for R to decrease from 100% to 90%."
• Difficulty is "the inherent complexity of the information."

The current version is FSRS-6, and it uses 21 parameters, optimized by machine learning against your own review log. (That count changed in Anki 25.06, so anything quoting an FSRS-5 parameter count is out of date.) It also gives you a dial SM-2 has no equivalent for: desired retention, settable from roughly 70% to 97%.

Those definitions are load-bearing in the table above. A stability of 6.55 days means the model put your odds of producing sacacorchos at 90% about six and a half days out, which is why failing on day 7 was informative rather than surprising.

FSRS was created by Jarrett Ye, a research engineer at MaiMemo, and developed by the Open Spaced Repetition community. It derives from MaiMemo's DHP model.

One honest note on the headline claim. The 20 to 30% fewer reviews for the same retention figure comes from FSRS's own docs. Treat it as directional.

What each app runs, checked 11 August 2026

App	Scheduler	Adapts to you?
Anki	SM-2 by default, FSRS-6 once enabled	Yes, fully, with FSRS on
RemNote	SM-2 by default, FSRS-6 as opt-in beta	Yes, fully, with FSRS on
SuperMemo	SM-19 in SuperMemo 19 (2024)	Yes
Clozemaster	Fixed mastery ladder, 0% to 100%	No, though Pro lets you edit the intervals
Memrise	Fixed ladder, 4h to 6 months	No
Duolingo	Published half-life regression (2016)	Unknown in production today
Quizlet	None documented	No due dates at all
Verbamor	FSRS	Yes

1. Anki

Anki is the right answer for most people who are serious about this, and I'd say so if I sold something twice as good.

It was released in 2006 by Damien Elmes, so it has close to 2 decades of development behind it, and the current desktop stable is 26.08.1, out on 5 August 2026. Desktop is free on Windows, macOS and Linux. AnkiDroid is free. AnkiWeb sync is free. iOS is the only paid platform: AnkiMobile is US$24.99 one time, and that purchase is what funds the whole project. Regional prices differ, so check your own store.

On scheduling: FSRS is opt-in, and SM-2 is the default. The manual calls FSRS "an alternative to Anki's legacy SuperMemo 2 (SM-2) algorithm" and tells you to "Enable FSRS under the FSRS section, at the bottom of the deck options page." I've verified that in both the manual and the Rust source, because a lot of SEO posts claim FSRS became the default in 23.10 and that is simply false. What happened in 23.10, October 2023, is that FSRS became built in, replacing the fsrs4anki add-on. Current Anki ships FSRS-6, with a hardcoded default desired retention of 0.9.

So if you installed Anki and never opened deck options, you're on a 1987 algorithm. Two minutes of setup gets you the best scheduler on this page for free.

The real trade is content. Anki ships no language content of its own. It's an empty engine, and you fill it with cards you build or decks the community shared. Building them yourself is a genuine advantage for retention: material you generate is remembered better than the same material read (Slamecka and Graf, 1978). It's also the reason a lot of people bounce off it in week 2. There are well over a thousand community add-ons if you want to reshape it.

2. RemNote

RemNote is the one to know about if the phrase "free FSRS" is what brought you here.

Its spaced repetition is free. The free tier lists "Unlimited Notes & Flashcards", and what the free tier caps is AI-generated cards (5 of them), rather than the SRS itself. Paid plans are Pro at $8 a month or $96 a year, and Pro with AI at $18 a month or $216 a year.

The scheduler story mirrors Anki almost exactly. The default is SM-2. FSRS is there, it's FSRS-6, and the docs are blunt about its status: "FSRS is a beta feature. It may become the default scheduler in the future, but for now you need to manually enable it." You'll find it under Settings, then Custom Schedulers.

Which means the 2 apps on this page with the most capable schedulers both ship them switched off. Any roundup that ranks apps by "uses FSRS" without checking the default is ranking a checkbox nobody ticked.

3. SuperMemo

Everything else here is downstream of SuperMemo. Wozniak's SM-2, published in 1987, is the algorithm Anki and RemNote both inherited, and the DSR ideas in FSRS grew out of the same research lineage.

The Windows line kept going. The history page lists SM-17 in 2015, SM-18 in 2019 and SM-19 in 2024, introduced in SuperMemo 19. If you read a comparison that stops at SM-17 or SM-18, it was written before 2024 or copied from something that was.

One practical caution. SuperMemo is split across several products, so "SuperMemo" on a price list may not mean the Windows program with the current algorithm. I'm not publishing a price because I couldn't verify one: the SuperMemo 19 listing on the official store showed as sold out when I checked, and a sold-out price is no use to you.

The most researched scheduler in existence, attached to the least approachable software here. I wouldn't start with it.

4. Clozemaster

Clozemaster drills vocabulary inside cloze sentences, which is a good format, and it's honest about being a fixed ladder if you read the docs.

The mastery levels are the schedule: 0% same day, 25% next day, 50% at 10 days, 75% at 30 days, 100% at 180 days. Four consecutive correct answers mark an item mastered. One wrong answer sends it back to 0%.

There's no per-card difficulty and no fitted model, so a word you keep fumbling sits on the same rungs as one you've never missed. Pro lets you customize the intervals manually. You pick the numbers, once, for everything.

The content is the draw. The Fluency Fast Track is 10,000 sentences in 10 levels of 1,000, ordered roughly by word frequency. There's a newer Fast Track and a Legacy version, and availability varies by language. I'm not publishing a price because I couldn't verify one.

5. Memrise

Memrise uses neither SM-2 nor FSRS. It runs a fixed, non-adaptive interval ladder, and the rungs are documented: 4 hours, 12 hours, 24 hours, 6 days, 12 days, 48 days, 96 days, 6 months. A wrong answer resets you to 4 hours.

Same schedule for every user and every item, so the word you can never hold onto and the word you got first try sit on identical rungs forever. That's the ladder worked through above, and the 12-to-48-day rung is where it costs the most.

Memrise has also moved on in its marketing. The homepage now leads with "Real native video. Smart review. Private AI practice," and the phrase "spaced repetition" doesn't appear on it. The ladder is still in the help centre. I'm not publishing a price because I couldn't verify one: it's region-varying and promo-heavy enough that any figure I quoted would be wrong for most readers.

6. Duolingo

Duolingo gets dismissed on this query, and the dismissal is wrong on the facts.

Duolingo published a real scheduler and open-sourced it: half-life regression, in Settles and Meeder, "A Trainable Spaced Repetition Model for Language Learning," ACL 2016. Peer-reviewed and trained on production data, which is more than Quizlet or Memrise has published.

What I can't tell you is whether HLR is what's running today. Secondary sources say it was folded into a later system around 2020, but I couldn't find that on a Duolingo-owned page, so I'm not going to state it. The published paper is real. The 2026 production model is undocumented.

Two other things worth correcting while I'm here. The free tier's gate is now Energy, a learning battery that lessons drain and correct answers refill, which replaced hearts. And the "Duolingo only gets you to A2" line is stale: since 22 April 2026 Duolingo says its Spanish course now teaches through CEFR B2, a Duolingo Score of 129. The expansion covers 9 courses. That's a claim about course content rather than measured learner outcomes, but repeating the A2 line in 2026 is just out of date.

Super Duolingo's App Store prices run from $9.99 to $119.99 depending on term and plan, with a 14-day free trial. The store doesn't label which price goes with which term, so I'm not going to guess.

7. Quizlet

Quizlet is on this list because people search for it here, and the honest finding is that it doesn't belong in the category.

I ran a full-text scan of all 164 English articles in Quizlet's help centre. Zero hits for "Memory Score", zero for "spaced repetition", zero for "scheduled review", zero for "forgetting curve." Learn is described as built on learning-science principles and powered by machine learning with a personalized study path. No named algorithm. No intervals.

What ships instead is Progress, and it has no time dimension. It buckets terms by how often you answered correctly (Not Studied, Still Learning, Mastered). No due dates and no next-review calculation. You decide when to come back.

One concession that's easy to miss: Learn and Test are paid for individuals, but students in a teacher-created class get unlimited ad-free Learn and Test on sets added to that class page. So "Learn is paid" is true for you and false for a lot of students. Quizlet's audio is machine text-to-speech, free in Flashcards mode, with custom recordings on the paid tier and native-speaker recordings not a feature at all. I'm not publishing a price because I couldn't verify one from a first-party page, so I'm describing what's gated instead. Roundups that quote you a flat Quizlet monthly figure are quoting each other.

8. Verbamor, and why it's last

Verbamor runs FSRS. So does Anki with 2 clicks, and so does RemNote's beta, on a free tier. On the algorithm, we're tied, and I'd be embarrassed to argue otherwise.

The narrowness is the point, and it's severe. iOS only. Four languages: Spanish, French, Italian and Portuguese. Paid, with no free tier at all: Explorer is $19.99 a month, Native is $129.99 a year (about $10.83 a month) with a 7-day trial on the annual plan. Anki does more than that for free on 5 platforms.

Verbamor isn't a general spaced repetition app. You can't put pharmacology in it, or Japanese, or your law-school outlines. If your deck isn't Spanish, French, Italian or Portuguese vocabulary, stop reading this section and go install Anki.

What it does is the step before the scheduler, and since this page has spent 2,000 words demanding other people show their working, here's mine. It records the lesson and reads the transcript, which arrives with speaker labels. It identifies you as the less fluent speaker: the one who hesitates, asks what a word means, reaches for an English word mid-sentence, or gets corrected. Then it tags each candidate item with a ranked reason. Struggled means you reached for it and couldn't produce it cleanly. Corrected means the other speaker fixed it after you got it wrong. Gap means you needed the word and never produced it at all, so it's inferred from what you talked around. Those 3 outrank merely taught or used.

You review the mined list and pick what to keep. Each survivor becomes a card carrying the sentence from your lesson with the answer blanked, plus 2 recordings of that sentence: one with the word replaced by a tone for the recall test, one with it spoken for the answer.

It misfires in ways worth knowing. Talk over your tutor and the speaker split degrades. Pause to think about a word you knew perfectly well and it can read as a struggle. None of that is a retention claim, because I don't have one. It's a mechanism you can judge on the trial. What the fixed-ladder apps can't match here isn't the scheduling, it's that they don't know your lesson happened at all.

How to pick

Three questions, in order.

Is your material anything other than these 4 languages? Anki. It's the only one here that takes any subject at any scale, and its scheduler is as good as it gets. Turn FSRS on.

Do you want FSRS without spending money on any platform? RemNote's free tier, with FSRS-6 enabled under Custom Schedulers, or Anki free on desktop and Android. Both give you the same algorithm class.

Is making the cards the reason you've quit twice? That's the one thing the free options don't solve, and it's the whole argument for a paid app in this category. Ours does it for 4 languages on iOS. Judge it on that, and only on that.

Whichever you land on, 2 things matter more than the choice: grade honestly, since a fitted model is only as good as the data you feed it, and mix your cards rather than blocking them by topic (see interleaving). If you want the scheduling principles without an app attached, the guide to spaced repetition for language learners covers the manual version, and the forgetting curve post explains why any of this works.

Questions people actually ask

---

Sources

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