Why We Threw Out the 1/3/7/21 Spaced Repetition Schedule [2026]

Search for a spaced repetition schedule and the same four numbers come back on every blog: 1, 3, 7, 21. Nobody can tell you who came up with them.

We went looking for the source because we had to pick intervals for LearnClash and wanted a citation to stand behind. There isn’t one. Sebastian Leitner’s 1972 boxes were measured in centimeters, not days. The algorithms inside Anki and SuperMemo compute every gap from recall history.

So we built a 3-stage system instead, and this piece is the argument for it: where the ladder came from (nowhere we could find), what Cepeda et al. (2008) actually measured, and the ladder we shipped and threw out ourselves before landing on 7 and 90 days. Duel me on memory psychology →

What Is the 1/3/7/21 Spaced Repetition Schedule?

It is a fixed-interval review method. You study new material, then review it on day 1, day 3, day 7, and day 21. Every card on the same ladder, every learner on the same calendar.

The 1/3/7/21 schedule as people imagine it: four fixed touch-points, every card the same.

You find it on Notion templates, Reddit study guides, and exam-prep posts on Medium. Clean, memorable, fixed. That is the entire reason it spread.

What the format never says out loud is that it treats every flashcard like every other flashcard. The vocabulary word you nailed on the first try gets the same day 3 review as the one you failed four times, and the kanji you have drilled for a month gets the same day 21 review as the one you met yesterday.

Real spaced repetition branches. Get a card right and the next review moves further out; get it wrong and the gap shrinks. A schedule that ignores how you are actually doing is not a learning system, whatever the template calls it.

Where Did the 1/3/7/21 Schedule Actually Come From?

Nobody knows. There is no founding paper, no journal citation, and no textbook chapter that introduces 1/3/7/21. Sebastian Leitner’s 1972 book So lernt man Lernen introduced the five-box system behind most modern SRS work, and his original was not a day calendar at all. It was five cardboard compartments measuring 1, 2, 5, 8, and 14 centimeters, and you reviewed a box when it filled up.

Leitner’s 1972 boxes beside the folklore that replaced them.

The trail for the ladder itself ends at blog posts citing other blog posts. Study YouTubers, then Notion templates, then productivity sites that rank for “best spaced repetition schedule” with no citation at all. We cannot tell you who wrote the numbers down first, and as far as we can find, neither can anyone who repeats them.

Even academic-adjacent sources do not help. Birmingham City University’s revision guide teaches a “2357 method” instead: study sessions spaced 2, 3, 5, and 7 days apart, counted backwards from the exam. Fixed again. And again, the page cites the forgetting curve in general but no paper that measured those particular numbers against alternatives.

To be fair to both ladders, they are not harmful. Reviewing on four spread-out days beats reviewing once, and spacing always beats massing. The problem is what they claim to be. They are presented as the intervals science found, and they are not; they are a rounded-off memory of a finding, repeated until people stopped checking.

What Does the Research Say About Fixed Intervals?

The science says the right gap depends on how long you need to remember. Cepeda, Vul, Rohrer, Wixted, and Pashler (2008) taught 32 trivia facts to 1,354 people, inserted a gap of up to 3.5 months before a single review session, and tested them up to a year later, 26 gap-by-delay combinations in all.

At every test delay, final recall rose as the gap grew, peaked, and then fell. And the peak moved. The best gap was about 20 percent of the test delay for tests a few weeks out, falling to about 5 percent when the test was a year away.

Figure 3: Cepeda’s temporal ridgeline. The optimal gap grows with the retention window, but more slowly than the window does, so the ratio keeps shrinking.

Here is the ridge in the paper’s own recall data:

Test delay	Best gap measured	Gap as a share of delay
7 days	1 day	14%
35 days	11 days	31%
70 days	21 days	30%
350 days	21 days	6%

Hold 1/3/7/21 against that table, remembering that the experiment measured one review per gap while the ladder stacks four. For a quiz a week out, the day 1 review sits right on the measured optimum. For a test a month out, the best gap was 11 days, so days 1, 3, and 7 all come too early and day 21 comes late. For a test a year out, day 21 is close to the best gap the paper measured, but it is also where the ladder stops, so the one review that fits is the last one you get. The ladder happens to land on the 1-day and the 21-day optimums and misses the 11-day one entirely, and no fixed ladder can fit all four delays at once.

Two things in the data matter more than the exact numbers. The ratio falls as the target grows, so no single ladder can be right for both a Friday quiz and a year-out exam. And the costs are lopsided: the authors note that a gap longer than the optimum costs you a little, while a gap that is too short costs you a lot. 1/3/7/21 errs short for everything past a few weeks.

One more study is worth knowing. Kang, Lindsey, Mozer, and Pashler (2014) taught foreign vocabulary over four weeks using either expanding or equal-interval retrieval practice, then tested eight weeks later. Final recall was equivalent. The shape of the ladder mattered far less than the fact that retrieval was spaced at all.

Robert Bjork’s lab calls the underlying principle desirable difficulties: training conditions that slow you down in the moment but pay off in long-term retention. Spacing is the textbook example, and the lab’s own framing is that it works best when the information is hard to access but not impossible. That zone sits in a different place for every card. A per-card scheduler can find it; a ladder cannot, which is why every serious SRS algorithm since 1987 has scheduled from recall instead of from a calendar.

How Do Real SRS Algorithms Schedule Reviews?

Piotr Wozniak’s SM-2, shipped in SuperMemo 1.0 in December 1987, was the first computer algorithm for computing a spaced repetition schedule. It keeps a per-item ease factor that stretches or shrinks the next gap depending on how you graded your own recall. SuperMemo kept iterating (SM-5 in 1989, SM-17 by 2016, SM-18 in 2019) and in 2026 reached SM-20, which SuperMemo describes as the first version where every parameter is fitted by machine learning.

Anki, released in 2006, built on SM-2 and still ships it as the default scheduler. Since version 23.10 in late 2023 it has also included FSRS as a built-in option you switch on per preset.

Four decades of scheduler design, 1987 to 2026. None of them produced 1/3/7/21.

Algorithm	Year	What it schedules on
Leitner box	1972	Manual promotion; a box is reviewed when it fills
SM-2 (Wozniak)	1987	Per-item ease factor and previous interval
Anki default scheduler	2006	SM-2 derivative with configurable learning steps
SuperMemo SM-18	2019	Two-component model: stability and retrievability
FSRS-6	current release	21 parameters fitted to difficulty, stability, retrievability
FSRS-7	in development	Fractional intervals; still research, not shipped
SuperMemo SM-20	2026	All parameters fitted by machine learning
LearnClash 3-stage SRS	2026	Accuracy gates at 7 and 90 days

Not one of them contains a fixed ladder. Every team that has sat down with review data and shipped a scheduler has rejected the idea, for the reason the Cepeda table makes obvious: a calendar cannot know what one learner remembers about one card.

FSRS is the interesting one because it is open. The project’s benchmark runs on roughly 727 million reviews from 10,000 Anki users. The current release, FSRS-6, fits 21 parameters to three per-card quantities (difficulty, stability, and retrievability) and its maintainers claim 20 to 30 percent fewer reviews than SM-2 for the same retention. The next interval is computed straight from that memory state, with no ladder to climb.

We did not try to out-model any of this. We picked a smaller, more visible state machine on purpose, and let it gate on per-card accuracy instead of a clock. That is the 3-stage SRS, and it took us three attempts.

Why We Built LearnClash’s 3-Stage SRS Instead

We did not arrive at three stages on the first try. The first memory system we shipped, on January 16, 2026, had six stages (Reset, Seen, Familiar, Known, Strong, Mastered) with cooldowns of 1, 2, 5, 14, and 45 days, plus a 60-day time gate before anything could count as mastered. On paper it looked rigorous. Looking back, it was our own fixed ladder with extra steps: five intervals we could not have defended from the research, and a 60-day clock running alongside the accuracy gates.

On February 3 we cut it to four stages with cooldowns of 1, 7, and 30 days. On March 20 we cut again, to the three stages that ship today: Learning, Known, and Mastered, with 7 and 90 days between the checks. The dates and the numbers come straight from our commit history. The intervals that survived are the two we can map onto Cepeda’s data; the ones we dropped, we could not.

Figure 5: the system as shipped. Promote on a correct answer at the check, drop one stage on a miss, leave the pool at Mastered.

A card’s first answer sorts it: right on first sight and it goes to Known, with the next check 90 days out; wrong and it sits in Learning and comes back in 7 days. Pass the 90-day check and the card is Mastered, which sets its next review so far in the future that it leaves the active pool. Miss at any check and the card drops exactly one stage. A Mastered card you miss goes back to Known, not to the start.

Two details that the diagram hides. If a card happens to come up again before its check (duels draw from a shared pool), a correct answer earns a small bonus but does not advance the stage and does not reset the clock. And “leaves the pool” has a footnote: a mastered card can still appear in a duel when a topic runs out of fresh material for both players, as last-resort filler capped per round. It is out of the review schedule, not banned from the game.

Why 7 and 90, then? Measured against Cepeda’s ridge, a 7-day first gap targets the month scale, where the paper found an 11-day optimum for a 35-day test, on a curve that falls off gently past its peak. The 90-day gap is the “delayed review of at least several months” the same paper recommends for anyone who wants to retain something for years. Both sit on the long side of their target, which is the cheap side to miss on.

We made three design calls that diverge from FSRS:

1. Coarse over fine. Three states beat thirty intervals when the user has to understand what is happening to a card without reading documentation.
2. A finite endpoint. Anki reviews cards forever; LearnClash retires them, so the active pool has a ceiling.
3. Demote one stage rather than reset. A miss is a signal, not a verdict.

That third call is a deliberate departure from SM-2, where repeatedly failing a graduated card can drive its ease so low that Anki’s manual has a name for it: ease hell. The honest cost on our side is real too. We bet that explainability is worth more than the squeezed percentage points, especially in a competitive app where the same player is also climbing a Glicko-2 rating and does not want to think about scheduler internals.

Which Apps Actually Run Performance-Based SRS in 2026?

Most of the serious flashcard apps schedule from per-card recall, and none of them ship a 1/3/7/21 ladder. The differences are in what the default is, and whether the schedule survives the session.

Seven apps, one question: does the next review come from your recall or from a calendar?

App	Scheduler	Per-card, across sessions	Note
Anki	SM-2 by default; FSRS built in since 23.10 (FSRS-6 today)	Yes	FSRS must be switched on
SuperMemo	SM-18 (2019), SM-20 (2026)	Yes	Closed source
RemNote	Anki SM-2 by default; FSRS available	Yes	FSRS must be switched on
Mochi	Own lengthen-or-shorten scheduler; FSRS available	Yes	FSRS is optional
Brainscape	Confidence-Based Repetition, you rate each card 1 to 5	Yes	Self-rated, not answer-checked
Quizlet	Learn mode	Within a session on the free sample; Learn itself is a Quizlet Plus feature	No scheduler published beyond a 2017 engineering post
LearnClash	3-stage SRS, checks at 7 and 90 days	Yes	On in every mode, no ads in any tier

Quizlet is the interesting case. Its Learn mode runs a logistic regression model trained on about 1.5 million sampled answers, and Quizlet’s own 2017 engineering write-up said 95 percent of users study a set over four days at most. So Learn ranks the terms you are closest to forgetting inside a session. As of Quizlet’s own help page (updated August 2026), Learn itself is a Quizlet Plus feature, and non-subscribers can start a free study session to see what it is like.

That is an architecture tuned for the use case Quizlet sees most (a weekend before a test), not for long-term memory. It is a choice, not a defect, but it is worth naming.

Our standing position: an app that ships 1/3/7/21 is running a calendar with extra steps. Show us the algorithm and the retention curve and we will change this article. For the longer Anki versus Quizlet comparison, including what FSRS does that Learn mode does not, we wrote that up separately.

If You Were About to Copy the Template

Take the spacing and leave the numbers. The 1/3/7/21 schedule is not from Leitner, not from Wozniak, and not inside any scheduler that ships in 2026. The closest thing to a source, Cepeda et al. (2008), says the right gap is a moving share of your retention target, and a single ladder cannot move. Pick a system whose scheduler responds to your recall, not to a date on a Notion template.

The verdict in one image.

If you want the rest of the argument, the learning science cluster covers retrieval practice and the LearnClash retention curve in more depth.

Sources used in this article:

• Cepeda, Vul, Rohrer, Wixted, and Pashler (2008), “Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention,” Psychological Science
• Kang, Lindsey, Mozer, and Pashler (2014), “Retrieval practice over the long term: Should spacing be expanding or equal-interval?” Psychonomic Bulletin & Review
• SRS Benchmark, open-spaced-repetition project
• Anki manual, deck options (FSRS, SM-2, ease)
• Sebastian Leitner system, So lernt man Lernen (1972)