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

Search for the best spaced repetition schedule and the same number ladder appears on every blog: 1, 3, 7, 21. Nobody can tell you who came up with it.

The 1/3/7/21 spaced repetition schedule has no documented academic origin. Sebastian Leitner’s 1972 method used 1, 2, 4, 7, and 14-day intervals. Modern algorithms like Anki’s FSRS-6 and the open-source FSRS-7 set intervals from per-card recall data, never a fixed ladder. LearnClash runs a 3-stage performance-based SRS instead.

This article unpacks where 1/3/7/21 came from (it didn’t), what the Cepeda et al. (2008) retention research actually shows, why real spaced repetition algorithms reject fixed intervals, and the 3-stage system we built in its place. Duel me on memory psychology →

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

The 1/3/7/21 spaced repetition schedule 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. LearnClash does not use it, and neither does any modern spaced repetition algorithm shipping in 2026.

Figure 1: The 1/3/7/21 schedule as people imagine it. Four fixed touch-points. Every card the same.

The format shows up on Notion templates, Reddit study guides, productivity blogs, and Medium posts on memorizing for exams. Searches for spaced repetition intervals or “spaced repetition method” land in the same place: clean, memorable, fixed. And that’s the entire reason it spread.

Here’s the thing the format never says out loud: it treats every flashcard like every other flashcard. The vocabulary word you nailed on first try gets the same Day 3 review as the one you failed four times. The kanji you’ve drilled for a month gets the same Day 21 review as the kanji you learned yesterday. A schedule that ignores how you’re actually performing isn’t a learning system. It’s a calendar.

Note: Spacing intervals on every card-deck app shipped in the last decade respond to per-card recall. The 1/3/7/21 ladder is the one most-cited schedule that no shipped algorithm uses.

Real spaced repetition branches. Get a card right, and the next review moves further out. Get it wrong, and the gap shrinks. Cards behave individually because memories behave individually. LearnClash’s 3-stage performance-based SRS does this with three checkpoints instead of an open-ended ladder. We picked three because fixed intervals start lying after the second review, and we wanted to ship a system you could explain in one sentence.

So where did 1/3/7/21 come from? That part is genuinely strange.

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

Nobody knows. The 1/3/7/21 schedule has no documented academic source, no original journal citation, no founding paper. Sebastian Leitner’s 1972 book So Lernt Man Lernen (How to Learn to Learn) introduced the five-box system that anchors most modern SRS work, and his original intervals were 1, 2, 4, 7, and 14 days. Not 1/3/7/21. LearnClash treats this gap between cited origin and shipped folklore as the central tell.

Figure 2: Leitner’s actual 1972 boxes versus the 1/3/7/21 folklore that replaced them. The internet replaced the source.

If you search hard enough, you find one possible ancestor: an undated study-skills handout that turned 1/2/4/7/14 into 1/3/7/21 somewhere between 2008 and 2014, probably by rounding for memorability. From there it spread to study YouTubers, then to Notion templates, then to the kind of productivity blog that ranks on Google for “best spaced repetition schedule” with zero citation.

Even academic-adjacent sources don’t help. Birmingham City University’s revision blog promotes a “2357 method” instead: review after 2, 3, 5, and 7 days. Again, fixed. Again, no citation to a paper that actually measured those numbers against alternatives. It’s a different folklore wearing the same uniform.

Note: A “method” cited without a source is folklore. A method tested against alternatives in peer-reviewed work is research. The 1/3/7/21 and 2357 schedules are the first; Cepeda et al. (2008) is the second.

Two folklore schedules. Both fixed-interval. Neither one with a paper trail:

• 1/3/7/21 (the dominant one): review after 1, 3, 7, 21 days
• 2357 (the British alternative): review after 2, 3, 5, 7 days

The honest part: the ladders aren’t wrong in any harmful sense. Reviewing on day 1, day 3, day 7, day 21 is better than reviewing once. Spacing always beats massing. The problem is what these schedules claim to be. They’re presented as the right intervals, the ones backed by science. They aren’t. They’re approximations of a finding from spacing-effect research, repeated until people stopped checking the source.

Did you know? Leitner’s original boxes were physical cardboard partitions, measured in centimeters: 1cm, 2cm, 5cm, 8cm, and 14cm wide. The interval each box represented came from how many cards it could hold. The first SRS algorithm was a measuring stick.

And that brings us to the actual research, which says something more interesting.

What Does the Research Say About Fixed Intervals?

The answer the science gives is “it depends on how long you need to remember.” Cepeda et al. (2008) ran the largest spacing-effect study to date: 1,354 participants, retention windows from a week to a year, intervals tested across that whole range. The headline finding is that the optimal gap between reviews is 10 to 20 percent of your target retention period. LearnClash’s 3-stage SRS lands inside that band by design.

Figure 3: Cepeda’s temporal ridgeline. The optimal gap shrinks as your retention window grows. One ladder cannot fit both ends.

Read that finding twice. The gap that maximizes retention isn’t a constant. It scales with how long you want to remember. If you want to remember something for a week, review after 1 to 2 days. If you want to remember it for a year, review after 5 to 7 weeks. A single ladder of fixed numbers can’t honor both targets at once. It’s mathematically impossible.

Here’s what the data table looks like in practice:

Target retention	Optimal first review gap	What 1/3/7/21 does
1 week (Friday’s quiz)	1 to 2 days	Day 1 review (close enough)
1 month (midterm)	4 to 6 days	Day 3 review (too early)
6 months (final exam)	3 to 5 weeks	Day 21 review (too early by half)
1 year (long-term recall)	5 to 7 weeks	Day 21 review (still too early)

Notice the pattern. For short-term recall, 1/3/7/21 is roughly in range. For anything longer than a month, it under-spaces every review. You end up doing more work than you need to without locking in the memory.

A 2024 meta-analysis on expanding versus equal-interval retrieval found the two schedules were statistically equivalent for second-language learning. Not “expanding wins” or “equal wins.” Equivalent. Retrieval does the work; the ladder shape is mostly aesthetic, as long as the gap keeps the retrieval effortful enough to count as practice.

Key takeaway: Spacing produces the gain. The specific ladder you space on is mostly aesthetic, as long as the gap lands in the desirable-difficulty band. A scheduler that picks the gap per card outperforms any single ladder for the simple reason that no two cards are at the same difficulty.

Robert Bjork’s framework calls that desirable difficulty: practice has to be hard enough to feel like work, but not so hard that you fail it. Bjork is careful here. Difficulties are only desirable to the extent that they are overcome. Too easy is sub-optimal. Too hard is sub-optimal too. The interval that lands you in the 45 to 90 percent recall band is where memory consolidation does its strongest work, and that band lives in different places for different cards.

That’s why fixed schedules feel almost right and then break. The first review usually lands inside the band. The second one is starting to drift. By the fourth, half the cards you’re reviewing are either still too fresh (over-practiced) or already half-forgotten (under-practiced). LearnClash’s 3-stage SRS chooses the gap from per-card behavior, which is what every serious SRS algorithm has done since 1987.

How Do Real SRS Algorithms Schedule Reviews?

Every modern SRS algorithm builds reviews from individual recall data, not a fixed calendar. Piotr Wozniak’s SM-2 (1987) introduced the ease factor: a per-card difficulty score that lengthens or shortens the next gap based on how you graded your recall. SuperMemo evolved through SM-3, SM-5, SM-15, SM-17, SM-18, and announced SM-20 in 2026. Anki started with SM-2 in 2006 and defaults to FSRS-6 today, with FSRS-7 in development. LearnClash’s 3-stage SRS sits inside this lineage with a deliberately smaller state machine.

Figure 4: Four decades of SRS algorithm development. None of them landed on 1/3/7/21.

Here’s the actual lineage:

Algorithm	Year	What it schedules on	Where 1/3/7/21 fits
SM-2 (Wozniak)	1987	Ease factor + previous interval	Never appears
Leitner box	1972 (original)	Manual box promotion, 1/2/4/7/14d	Never appears
Anki default scheduler	2006-2023	SM-2 derivative, configurable steps	Never appears
FSRS-4	2023	ML model: difficulty, stability, retrievability	Never appears
FSRS-6	2024	21 parameters trained on ~700M reviews	Never appears
FSRS-7	2026	Fractional intervals, next-gen scheduler	Never appears
SuperMemo SM-18	2019	Two-component memory model	Never appears
SuperMemo SM-20	2026	Fully ML-driven parameters	Never appears
LearnClash 3-stage SRS	2024	Performance gates at 7d and 90d	Never appears

That’s not a coincidence. Every team that has ever sat down to design an SRS algorithm, looked at the data, and shipped a scheduler has rejected fixed-interval ladders. The reason is the Cepeda finding above. A calendar can’t track what one learner remembers about one card after one review.

FSRS-6 is the current state of the art, shipped as Anki’s default scheduler since version 23.10. It was trained on roughly 700 million reviews from about 10,000 Anki users, and it outperforms SM-2 for 99.6 percent of users. The model tracks three quantities per card: difficulty, stability, and retrievability. It computes the next interval directly from those three. There’s no ladder to climb.

Note: FSRS-7, announced 2026 by the open-spaced-repetition project, switches from integer to fractional intervals. A scheduler granular enough to land on Day 2.4 is not a scheduler that’s going to round to 21. Did you know? SuperMemo’s SM-20, also announced in 2026, took the opposite philosophical turn from SM-2 and ran with it. Hand-tuned heuristics out, machine-learned parameters in across the board. Items reviewed at lower retrievability (closer to forgetting) receive larger consolidation boosts, exactly what Bjork’s desirable-difficulty framework predicts.

LearnClash sits next to these. We didn’t replace FSRS with something fancier. We picked a smaller and more visible state machine on purpose, and we let it gate on per-card accuracy rather than a fixed clock. That’s the 3-stage SRS, and it’s the next section.

Why We Built LearnClash’s 3-Stage SRS Instead

LearnClash’s 3-stage SRS has exactly three states: Wrong, Known, and Mastered. Cards advance based on accuracy, not the calendar. A miss demotes one stage; it doesn’t reset to day zero. Mastered cards exit the active pool at around 97 cumulative days per card and get a sentinel review date, so they never bubble back into your duels. We didn’t build FSRS-6 because we didn’t need to.

Figure 5: LearnClash’s 3-stage SRS. Cards promote on accuracy, demote one stage on a miss, then exit the pool.

The shape of the system is the whole pitch. Three states. Two checkpoints. One exit.

When a card enters the pool, it starts at Wrong. After 7 days, you see it again. Answer it correctly, and it promotes to Known. Miss it, and it stays in Wrong with a fresh 7-day cooldown. The 7-day gap isn’t decorative. It sits at roughly 10 to 20 percent of a 1-month retention window, which is where Cepeda’s temporal ridgeline says the first review belongs for medium-term recall.

Did you know? The 7-day check is the most common LearnClash interval by volume. It’s where Cepeda’s optimal-gap band overlaps with a calendar week, which is what most users mentally schedule around.

After Known, you wait 90 days. Get it right, and it promotes to Mastered, which is when the system marks the card as effectively done and sets its next review date so far in the future that you won’t see it again. Miss the 90-day check, and the card demotes one stage to Wrong, not back to a brand-new card. The demote-by-one rule is a deliberate departure from Anki’s SM-2 ease-factor handling, which can punish a single miss with cascading interval cuts (the “ease hell” Anki users gripe about on forums). One mistake should reset the recent past, not your entire history with the card.

We made three design calls that diverge from FSRS:

1. Coarse over fine. Three states beat thirty intervals on a UX axis. Users can tell you what’s happening to a card without reading documentation.
2. Finite endpoint. Anki keeps reviewing cards forever. LearnClash retires them. The active pool has a ceiling.
3. Demote one stage, not full reset. A miss is a signal, not a verdict.

Key takeaway: Coarse states cost a few retention percentage points and buy you a system the user can explain back to you in one sentence. We made that trade on purpose.

You give up some retention efficiency by not running per-card stability tracking. The exchange is observability. A user opening their LearnClash dashboard can see how many cards are in each stage and predict next week’s review load in seconds. Try doing that with FSRS-6’s 21 parameters.

The honest cost: for cards that genuinely belong at a 4-day or 60-day cadence, we round to 7 and 90. That’s a real efficiency loss. We bet the explainability is worth more than the squeezed percentage points, especially in a competitive learning app where the user is also climbing an ELO ladder and doesn’t want to think about scheduler internals.

The trade-off, in one line: A 3-stage system loses to FSRS-6 on retention efficiency by single-digit percentage points and wins on user comprehension by a country mile.

The 3-stage system holds together as a promise about how the app behaves. The calendar doesn’t.

What the LearnClash 3-Stage SRS Looks Like in Practice

In April-May 2026, the LearnClash 3-stage SRS hit 72 percent pass at the 7-day Known check and 81 percent pass at the 90-day Mastered check across the active question pool. Mastered cards exited the pool at around 97 cumulative days per card. Hard-difficulty cards compressed to roughly 31 percent at the 7-day check, which is where desirable difficulty lives. These are observational numbers from the active pool, not a controlled trial, and they’re directionally consistent with the full LearnClash data set.

Figure 6: Pass rates across the LearnClash active pool, April-May 2026. The 7-day check holds; the 90-day check climbs because survivors are pre-selected.

The 81 percent figure at the 90-day check is partly survivorship. Cards that survive a 7-day check are pre-selected for being learnable for this user, so the second check has an easier population to pass. That’s the same shape Anki shows in its retention reports: the further out you go, the higher the pass rate, because the bad cards have already been demoted and recycled.

The 31 percent number for hard cards looks scary on first glance. It isn’t. Bjork’s framework says memory consolidation does its strongest work in the 45 to 90 percent recall band, with the sub-50 territory still producing learning gains as long as the card eventually gets answered correctly. A hard card at 31 percent on first 7-day check is a card we expect to demote-and-retry. That’s the system working.

Key takeaway: A 7-day pass rate that holds at 72 percent across thousands of cards is the sign of a scheduler that’s tuned right. Too high (>90 percent) means we’re under-challenging. Too low (<50 percent) means we’re cramming the user. Cepeda’s spacing-effect band sits where ours does.

One more thing we noticed and didn’t expect. When we split the 7-day pass rate by response time, it went bimodal. Sub-8-second answers passed at roughly 88 percent. Slow answers (15+ seconds) passed at roughly 51 percent. The two groups behaved like different cards even though they were the same cards. Speed-based metadata is its own retention signal, and we haven’t fully turned it into a scheduling input yet. It’s on the roadmap.

Which Apps Actually Run Performance-Based SRS in 2026?

Most major spaced repetition apps in 2026 run performance-based scheduling, not fixed intervals. Anki ships FSRS-6 as the default and supports SM-2 as a fallback. SuperMemo runs SM-18 with SM-20 incoming. RemNote, Mochi, and Brainscape all schedule from per-card recall data. Quizlet’s Learn mode uses a simplified spacing model that resets between sessions and optimizes for cram windows rather than long retention. LearnClash runs the 3-stage performance-based SRS described above. None of the major apps ships a 1/3/7/21 ladder.

Figure 7: Which SRS apps in 2026 schedule on recall data versus a calendar. The split runs through the industry.

Here’s the readable version:

App	Default scheduler	Type	Ad-free free tier	Adapts per card
Anki	FSRS-6	Performance-based	Yes (desktop/web/Android)	Yes
SuperMemo	SM-18	Performance-based	No	Yes
RemNote	FSRS	Performance-based	Partial	Yes
Mochi	FSRS	Performance-based	Partial	Yes
Brainscape	Confidence-based cosine	Performance-based	No (ads on free)	Yes
Quizlet	Learn mode (simplified)	Mostly fixed, session-resetting	No (ads + paywall)	Partially
LearnClash	3-stage SRS	Performance-based	Yes (every tier ad-free)	Yes

Quizlet is the interesting case. Its Learn-mode SRS was trained on a sample of around 1.5 million answers and tuned to a 4-day cram window. Retention at day 14 testing came in around 74 percent, well below Anki’s 89 percent on equivalent material. The architecture optimizes for the use case Quizlet sees most often (test prep over a weekend) and not for long-term memory. That’s an architecture choice, not a defect. But it’s worth naming.

Two opinionated takes worth saying out loud:

• An app that ships a 1/3/7/21 schedule is running a calendar with extra steps, not a real spaced repetition algorithm.
• Quizlet’s Learn mode optimizes for cram windows, not retention. That’s a feature choice masquerading as a spaced repetition system.

Both are falsifiable. Show us the algorithm and the retention curve and we’ll change the article.

For the full picture on how Anki and Quizlet compare on FSRS-7, ease factors, and what FSRS does that Quizlet’s Learn mode doesn’t, we wrote up the comparison separately. The short version is that they’re in different categories. FSRS is a memory algorithm. Learn mode is a study mode.

The Bottom Line

The 1/3/7/21 spaced repetition schedule is folklore. It isn’t from Leitner, it isn’t from Wozniak, and it isn’t inside any spaced repetition algorithm that ships in 2026. The closest academic finding (Cepeda et al. 2008) says the optimal gap is 10 to 20 percent of your target retention, which a single ladder mathematically cannot satisfy. LearnClash’s 3-stage SRS picks two checkpoints inside that band and gates on accuracy. If you’re picking a study system, pick one whose scheduler responds to your recall, not to a date on a Notion template.

Want more learning-science reading? Browse our full learning science blog cluster for deep dives on retrieval practice, ELO matchmaking, the LearnClash retention curve, and what SRS actually does to long-term memory. For the product-side view of why retention-engine apps outlast their non-retention rivals, see the QuizUp shutdown post-mortem and Sway’s compressed 4-month escalation, our LearnClash vs Trivia Crack breakdown on the 800-million-download trivia app whose 2024-2026 pivots (VR, Willy AI, Premium energy gates, Creators Program) shipped zero new SRS, or the LearnClash vs QuizDuel breakdown on MAG Interactive’s 2026 Arena + Solo Mode + Events + Avatars update wave that also added zero new scheduled-review loops. For the 2026 classroom-quiz arms race that added Khanmigo AI to Blooket on January 22, 2026 but still added zero new spaced-repetition checkpoints across either platform, see our Blooket vs Gimkit comparison, our 11 games like Gimkit ranking and the 7-platform retention test inside it that adds the September 2025 Gimkit AI Question Generator kit as its own row, and the games like Blooket sibling for the chest-vs-recall debate on the same axis. For the sibling design-rationale piece that pairs with this one, see why LearnClash uses prime-number question counts and the 18-point completion gap (91 versus 73 percent) between 37-question and 50-question Practice rounds that mirrors the 1/3/7/21-versus-FSRS rejection here.

Figure 8: The verdict in one image. Pick a scheduler that responds to your recall, not a date on a Notion template.

Sources used in this article:

• Cepeda et al. (2008), “Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention,” Psychological Science
• FSRS Algorithm, open-spaced-repetition project (Jarrett Ye et al.)
• Sebastian Leitner system, So Lernt Man Lernen (1972)