Anki is a spaced-repetition engine you point at cards you build yourself. A modern SRS app is a language product that schedules reviews on your behalf and never shows you the algorithm. The distinction that actually matters is not SM-2 versus FSRS, because Anki now supports both. It is whether you want control over every card and every setting, or want scheduling to disappear into a product that also supplies the content you are reviewing.
What a spaced repetition system actually decides#
Every SRS answers one question: given what I know about this item and this learner, how many days should pass before showing it again? The target is the moment just before forgetting, which is where retrieval is effortful and therefore most productive. The underlying evidence is old and unusually solid: Hermann Ebbinghaus mapped the forgetting curve in 1885, and a large meta-analysis of distributed practice studies (Cepeda et al., 2006) confirmed that spacing reviews beats massing them across a wide range of materials and intervals. Spaced repetition for language learning covers that research in full.
What differs between systems is only how the next interval is computed and how much of that computation you are allowed to see or change.
SM-2: the formula still under most apps#
SM-2 came out of SuperMemo, the program Piotr Woźniak began building in the late 1980s; the algorithm was described in his 1990 write-up and has been reimplemented everywhere since. Each card carries an ease factor that starts at 2.5 and is floored at 1.3. The first successful interval is one day, the second is six, and every interval after that is the previous one multiplied by the current ease. Grade a card badly and its ease drops, so the multiplier shrinks and the card comes back sooner.
Its strength is that it is transparent and cheap. You can compute the next due date on paper. Its weakness is that the formula was hand-tuned on one person's data and treats every learner's memory as obeying the same curve. Ease factor also does double duty as both a difficulty estimate and an interval multiplier, which is why badly graded cards in classic SM-2 can spiral into "ease hell," returning far more often than they need to.
FSRS: what actually changed in the 2020s#
FSRS, the Free Spaced Repetition Scheduler, is an open-source project developed by Jarrett Ye with a group of contributors during the early 2020s. It replaces the single ease number with three separate quantities: difficulty (how hard this item is for you), stability (how long the memory currently lasts), and retrievability (the probability you would recall it right now). That DSR framing descends from the same line of thinking as Bjork and Bjork's distinction between storage strength and retrieval strength (Bjork & Bjork, 1992).
The practical change is that FSRS does not use fixed constants. It fits its parameters to your own review log, so a scheduler trained on your history behaves differently from one trained on someone else's. You also set a desired retention level directly, choosing the trade-off between workload and recall rate instead of inferring it from an ease factor. For learners with thousands of reviews on record, that usually means fewer reviews for the same retention.
Anki today runs both engines#
Anki, released by Damien Elmes in 2006, has long used a modified SM-2: classic intervals plus learning steps, four grading buttons, leech detection, and per-deck interval modifiers. Since Anki 23.10, released in October 2023, FSRS ships in the app as an optional scheduler you enable in deck options and then optimize against your review history. Subsequent releases have refined it further.
This matters for comparison writing, and it is where most "Anki is outdated" arguments fall over. Anki is not stuck in 1990. It is currently the most configurable SRS available to an ordinary learner, and it can run the newer algorithm as well as the older one. What it does not do is decide what your cards should contain.
How modern apps embed the scheduler#
Newer language apps take the opposite approach: the scheduler is an implementation detail. Duolingo published its own model, half-life regression, which predicts recall probability from a learner's history and was described by Settles and Meeder in 2016. Vocabulary apps like Memrise, Clozemaster, and Drops each maintain review queues you never configure. In these products the algorithm is not a feature you tune; it is plumbing behind a lesson you tap through.
LingoBlend sits in that second camp deliberately. Its five practice games run on Anki-style SM-2 with sub-day learning steps of 10 minutes, 1 hour, and 8 hours before a card graduates to day-scale intervals, with ease bounded between 1.3 and 2.5. The reason for hiding it is that the cards come from your own reading, so the harder problem is card creation, not scheduling. Tapping a word inside a blended text creates the card, and the queue takes it from there.
Algorithm comparison#
| SM-2 (classic) | Anki with FSRS | Embedded schedulers | |
|---|---|---|---|
| Origin | Woźniak, SuperMemo, 1990 | Open source, early 2020s; in Anki since 23.10 (Oct 2023) | Proprietary, per app |
| Memory model | One ease factor per card | Difficulty, stability, retrievability | Varies; often undocumented |
| Parameters | Fixed constants | Fitted to your review history | Fixed by the vendor |
| You can tune it | Ease and interval modifiers | Desired retention plus optimization | Usually not at all |
| Grading input | Quality rating per review | Four buttons per review | Often just correct or wrong |
| Card creation | You write every card | You write every card | Generated from app content |
| Failure mode | Ease hell on hard cards | Needs review volume to fit well | You cannot fix what you cannot see |
Who Anki is genuinely better for#
Anki is the better choice if you want ownership. It is free on desktop and Android, stores your collection locally, syncs across devices, imports and exports without lock-in, and has an add-on ecosystem that covers nearly any workflow you can describe. The shared-deck library is enormous, sentence mining pipelines are built on it, and FSRS optimization gives long-term users measurable review savings. If you are studying for a fixed exam, or you already maintain a deck you trust, nothing on this list replaces it. Guidance on moving existing decks between tools is in importing Anki decks and word lists.
Anki's cost is time. A good card takes thought, and most people who quit Anki quit because the deck stopped being maintained, not because the scheduler failed them. That is the specific problem a context-first app is trying to remove, and Anki alternatives that keep the context compares the options that do.
Where an embedded scheduler wins#
Embedded scheduling wins when the bottleneck is supply. If you read an article and tap six unfamiliar words, six cards exist with their sentence, their base form, and their grammar already attached, and they enter a queue without you opening a card editor. The scheduling quality is probably slightly worse than a well-optimized FSRS setup. The number of cards that actually get made is far higher, and a mediocre schedule applied to words you met in real reading beats a perfect schedule applied to an empty deck.
The honest summary is that these tools compete on different axes. Anki optimizes the review. Embedded systems optimize the capture. If you want to see how the capture side works before comparing further, the features overview lists what each tier includes.