The LearnClash SRS Retention Curve: 3 Stages [2026]

The 1/3/7/21 schedule on every memory blog is a calendar dressed up as spaced repetition.

LearnClash’s 3-stage SRS runs on performance instead. A card you miss returns after 7 days. A card you know returns after 90 days. Clear both checks and the card retires from the review schedule. The fastest a card can manage that is 90 cumulative days (97 if it started with a miss), and most take longer, because a checkpoint only fires when the card actually comes back around in play. That three-checkpoint curve is the artifact this article describes, including its unglamorous corners: what “retired” literally means in the code, and the one loop a card can get stuck in. Duel me on memory psychology →

Below: how the curve is shaped, why we picked 3 stages over an interval ladder, what demote-by-one actually does, and how the design treats cards at the edge of recall.

What Is an SRS Retention Curve?

An SRS retention curve plots the percentage of cards a learner still answers correctly at each spaced interval. In LearnClash, the curve traces three checkpoints: how many cards pass the 7-day check, how many pass the 90-day check, and how many retire as mastered. It’s our overlay on the forgetting curve Hermann Ebbinghaus measured in 1885.

The raw forgetting curve falls to about a third within a day. The 3-stage SRS resets the slope at day 7 and day 90, then stops scheduling the card.

The unmodified forgetting curve is brutal. On Ebbinghaus’s relearning-savings measure, 58% survived at 20 minutes, 44% at one hour, 33% at 24 hours, and 21% at one month. Murre and Dros replicated the experiment in 2015 with Dutch syllables instead of German ones and got nearly the same shape, 130 years later. Ebbinghaus ran the whole 1885 study on himself, alone in a room with nonsense syllables, and the dozens of replications since have failed to bend his curve. It’s one of the most durable findings in cognitive psychology.

A retention curve is what you get when you draw the opposite of forgetting: each spaced review resets the slope. The interval-by-interval pass rate is how a learning app can check whether its SRS is actually doing its job.

LearnClash’s curve has three points instead of an open-ended ladder. That’s the design choice the rest of this article unpacks.

What Shape Does the LearnClash Retention Curve Take?

The curve traces a 7-day Known check, a 90-day Mastered check, and a retirement point where a twice-cleared card leaves the review schedule. The funnel between those checkpoints is measurable in production. As of the July 21, 2026 census, the SRS ledger held 128,085 card-player pairs across 1,161 accounts: 41.0 percent sat in the Wrong loop waiting on a 7-day check, 57.3 percent in the Known loop waiting on 90 days, and 1.7 percent had cleared both and retired. Source: LearnClash production export, July 21, 2026. That thin Mastered slice is a 90-day gate doing its job on a catalog where most cards are younger than 90 days.

Figure 2: The shape of the curve. The two checkpoints reset the forgetting slope; mastery ends the schedule. Random review, with no SRS, flatlines inside the same window.

Two things shape that curve, and both come from the learning literature rather than the calendar.

The 7-day check sits well above what a raw forgetting curve predicts at the same interval (Ebbinghaus’s savings ran between 21 and 33 percent for nonsense material at comparable delays; meaningful material fares better). The lift comes from active recall during the original encounter. In Roediger and Karpicke’s 2006 experiments, students who were tested on a prose passage reproduced 61 percent of it a week later, while students who re-read it managed 40 percent.

The 90-day check behaves counterintuitively: the cohort arithmetic favors it. A card that reaches the 90-day check has already cleared the 7-day check, so the 90-day cohort is pre-selected for cards the player actually encoded. Survivor bias works in the curve’s favor here, and against it on hard cards, as we’ll see further down.

Spacing carries the rest of the load. Latimier, Peyre, and Ramus pooled 29 studies in 2021 and found distributed retrieval practice beat massed practice by about three quarters of a standard deviation of retention.

Why 3 Stages Beat the 1/3/7/21 Heuristic Most Memory Blogs Cite

The popular 1/3/7/21 schedule reviews everyone on the same four calendar days regardless of what they remember. LearnClash’s 3-stage SRS moves a card forward only when the player gets it right; elapsed time on its own moves nothing. We chose 3 stages because the data calls for a few performance gates, not a daily cadence borrowed from a study planner.

Figure 3: The 1/3/7/21 ladder schedules everyone identically. The 3-stage SRS branches on whether the player actually recalled the card.

Here’s the side-by-side:

	1/3/7/21 heuristic	LearnClash 3-stage SRS
Trigger for next review	Calendar day	Player accuracy
Stages	4 (or more)	3 (Wrong, Known, Mastered)
Wrong-answer handling	Reset to day 1	Demote one stage
Mastery	Open-ended	Retired after two clean passes
Backed by	Study-planner blogs	Per-card recall + the Cepeda spacing dataset

The 1/3/7/21 idea has a real cognitive-science cousin. Cepeda and colleagues (2008) taught facts to more than 1,350 people, varied the review gap from zero to 105 days, and tested up to a year out. What they found was scaling, not a schedule: the optimal gap is a fraction of how long you want to remember, sliding from roughly 20 to 40 percent of the retention interval for week-scale goals down to 5 to 10 percent for year-scale goals. Rohrer and Pashler’s 2007 summary of the same research program gives the rule of thumb most people quote: space your reviews at about 10 to 20 percent of the time you want to keep the material.

The 10 to 20 percent rule of thumb visualized. It’s Rohrer and Pashler’s compression of companion Cepeda-led experiments; the underlying band actually slides from 20-40% for short goals to 5-10% for year-long ones.

For a one-week test, that means reviewing after a day or two. For retention measured in years, after one to a few months. No single ladder fits every goal.

Retention goal	Optimal-gap band (10-20% rule)	LearnClash interval
1 week	1-2 days	n/a (covered by the first encounter)
1-2 months	3-12 days	7-day Wrong→Known check
1-2 years	36-146 days	90-day Known→Mastered check

To be straight about the fit: 7 and 90 are calendar-round numbers, and they land inside or near the band rather than dead center. Seven days slightly overshoots a strict one-month optimum (3 to 6 days) because it’s tuned for a one-to-two-month horizon. Ninety days sits comfortably in the band for year-scale retention. Two checkpoints, two genuinely different retention windows, and no pretense that one ladder covers everything from Friday’s quiz to permanent knowledge.

So why three stages and not five or twenty? Two reasons.

The statistical reason

With two interval checkpoints, you need exactly three states: pre-check, between-checks, post-check. Every additional stage adds reviews for every player on every card, while the scheduling information a new checkpoint buys keeps shrinking once a card has already survived a 90-day gap. More stages means more reviews without much retention payoff.

The product reason

Players don’t want a card cycling forever. The pure-SRS school keeps reviewing into the indefinite future; LearnClash retires cards from the schedule. Mastered means done, not “we’ll see you in 365 days.” The longer-form case for why LearnClash threw out the 1/3/7/21 schedule walks through the missing source and the Leitner receipts.

How Does Demote-by-One Differ from Anki’s Reset?

A wrong answer demotes the card by exactly one SRS stage. A missed Known card drops to Wrong and waits 7 days. A missed Mastered card drops to Known and waits 90. Nothing ever resets to day zero. That’s a deliberate departure from how Anki’s classic scheduler treats misses, and it matters more than people expect.

Figure 5: Three algorithms, one miss. Demote-by-one sits between SM-2’s ease penalty and FSRS’s full model re-estimate.

Here’s how the three algorithms handle a wrong answer:

Algorithm	Wrong answer effect	Long-term consequence
Anki SM-2	Ease drops 20 percentage points, card relearns	”Ease hell” if a card flips repeatedly
Anki FSRS	Difficulty/stability/retrievability re-estimated	Marginal per-review adjustment
LearnClash 3-stage SRS	Demote one stage; that stage’s cooldown restarts	No ease drift; three inspectable states

SM-2 has a failure mode the Anki community named ease hell: cards that repeatedly lose ease pile up into a near-daily grind that never clears. The open-spaced-repetition team built FSRS partly to fix that. By modeling difficulty, stability, and retrievability as separate quantities instead of one ease number, FSRS reaches the same retention with about 20 to 30 percent fewer reviews, and the public benchmark behind it shows it out-predicting SM-2 for nearly every one of the roughly ten thousand user collections tested. It’s still opt-in, though, and the switch is collection-wide; Anki vs Quizlet covers the FSRS-6 numbers and the February 2026 AnkiHub stewardship handoff. For the classroom side of the same comparison, Kahoot vs Quizlet maps the forgetting curve against Kahoot’s Accuracy Mode, which has no SRS at all, and Quizlet Learn, whose spacing resets between sessions.

Our demote-by-one sidesteps both failure modes by carrying less machinery. There’s no ease factor to drift. There’s no per-card retrievability model to mistune. A card is in exactly one of three states, and a wrong answer moves it back exactly one. That makes the algorithm trivial to reason about and trivial to debug, and it means a player can always tell what state a card is in and why.

The trade-off is gentler punishment for a miss. SM-2 can swing a lapsed card into a daily relearning loop until it sticks; we make a missed card wait a full week before it’s scheduled again. Too long if you’re cramming for Friday’s exam. Intentional for the lifelong-learning loop a duel turn you can finish in a few minutes is built around.

Key takeaway: Demote-by-one trades fine-grained ease control for transparency. It loses nothing to ease hell because there is no ease to lose.

What Does the 90-Day Mastery Threshold Actually Mean?

A card is mastered after two off-cooldown correct answers: one to reach Known, one more, at least 90 days later, to reach Mastered. Then it leaves the review schedule.

My favorite implementation detail in the whole system lives here. Mastered isn’t a boolean flag; in duelConstants.ts the constant is SRS_MASTERED_DATE_MS, a next-review date of December 31, 9999. Every card sorts on the same next-review field, and a mastered card is simply one whose review is scheduled 7,973 years out. I find that more honest than a flag, because it states the definition directly: mastery means no review on any calendar you’ll live to see.

A card’s path from first encounter to mastery. A never-missed card can retire in 90 days; one that started with a miss needs at least 97; real cards take longer, since checks only fire when the card next comes up in play.

The fine print, straight from the same file: mastered cards can still show up. When a topic’s fresh supply runs short, the selector re-serves mastered cards as filler, capped per round. Those re-encounters advance nothing, and a miss on one demotes the card back to Known with a real 90-day cooldown. So “retired” means retired from scheduling, not banished from the game.

The 90-day number itself is the back half of the spacing band above: 10 to 20 percent of a retention goal measured in years. A card that survives a 90-day gap has consolidated well beyond session memory. That’s the territory a 2025 Cell Reports study by Zou, Kuhl, DuBrow, and Hutchinson maps at the neural level: tracking memories for images seen at intervals from seconds to months, they found the benefit of spacing was predicted by how the ventromedial prefrontal cortex re-encoded the earlier encounter.

Figure 7: The consolidation story behind the 90-day gap. Surviving it is the signal we retire the card on.

Why retire mastered cards at all? Three reasons.

Reason	Why it matters in LearnClash
Cognitive load	A cleared card burns 1 of 18 duel-question slots; that costs far more than Anki’s few-second review
Player motivation	Mastery has to mean done, not “see you in 5 years”
Pool freshness	Mastered cards leave the schedule, and topic catalogs stay demand-sized (median 27 questions per topic, July 2026 census)

The cognitive-load argument is the easy one. A card a player has cleared at 90 days is no longer informative for the scheduler, and re-asking it burns a duel slot that could go to a card the player still needs. Anki can afford to review forever because a 5-year-out review costs a few seconds of one person’s time; a duel slot is shared, scarce, and competitive.

The motivation argument is subtler. If a “mastered” card comes back five years later, mastered was just a label. The sentinel date ties the word to a real terminal state.

The pool-freshness argument is the one most pure-SRS apps ignore. Mastered cards drop out of a player’s active rotation, and the catalog behind that rotation is demand-sized rather than bloated: per the July 21, 2026 census, the median topic holds 27 questions. That’s a catalog-depth figure, not a per-player card count, and it’s the reason a lean active pool plus a bank of mastered cards beats a 200-card review backlog, which is just slower, not more rigorous.

How Does the Retention Curve Hold Up Across Difficulty Tiers?

The shape of the retention curve holds across difficulty bands, but the height drops with difficulty. Easy questions clear the 7-day check far more often than hard ones, and that’s by design. The difficulty bands track the Trivia Difficulty Index, our wrong-rate ranking across 570 topics, and they sit where desirable-difficulty research predicts they should.

Retention by difficulty band. The hard line stays low and flat across re-encounters; the easy line floats high throughout.

A first reaction is “fix the hard questions, the pass rate is too low.” That instinct points the wrong way: a low hard-tier pass rate is what the desirable-difficulty framework predicts; Bjork’s account has been tested directly in spacing-and-retrieval experiments like Maddox and Balota (2015), and Pyc and Rawson’s retrieval effort hypothesis found that harder successful recalls produce more durable memory than easy ones.

Figure 9: The desirable-difficulty zone. Recall right at the boundary, where retrieval only just succeeds, is where the retention lift is largest.

What does a question at the recall boundary look like? Take one of the hard-band items in our Life of Jesus Christ topic: it asks for the final words of Jesus as the Gospel of John reports them, while the decoys are genuine crucifixion sayings from the other gospels. Of twelve recorded answers, eleven went to a wrong gospel’s line. That’s not a broken question. It’s a trap built from adjacent truth, which is exactly what keeps hard-tier pass rates low and hard-tier learning high.

The flatness of the hard-tier line is also expected. Easy cards stay high across all checkpoints because recall is automatic. Hard cards stay low because each successful recall barely strengthens the trace; the cognitive effort on encounter two looks a lot like encounter one. That flat line is the curve’s signature for a difficulty band sitting at the edge of recall.

The practical implication for how to study effectively: if your goal is durable knowledge, don’t grind the same hard cards harder. Add topics, and let the boundary questions consolidate across many spaced encounters. The companion production list, 12 questions people get wrong, shows the recall edge in the wild: bananas as berries, Venus phases, skunk warning behavior, all misses that come back stronger after the reveal.

How the Curve Handles Cards at the Recall Edge

The shape of the 3-stage curve raises three design considerations worth calling out plainly: response speed carries information the final answer hides, hard cards stay flat rather than climbing, and demote-by-one allows a “stuck card” edge case. Each one shapes how the next iteration of the algorithm could evolve.

Three design considerations at the recall edge: response speed as a hidden signal, the flat hard-tier line, and the Wrong-Known cycle.

Consideration 1: response speed carries information the final answer hides

A fast, confident recall and a slow, uncertain one both count as “correct,” but they sit at different points on the memory curve. The slow one is where the fluency illusion lives: it feels like recall while slow recognition reconstructs the answer from partial cues. We deliberately never score speed (the 45-second timer exists; the points don’t care), so today both answers advance a card the same way. As a scheduling signal rather than a scoring one, response time is the most obvious candidate for a future version of the algorithm: it could separate genuine recall from reconstruction without ever touching the scoreboard.

Consideration 2: hard-tier retention stays flat, it doesn’t climb

You might expect the hard-tier line to climb across re-encounters, since each successful recall should strengthen the trace. The same survivor bias from earlier, now working against the curve, predicts it won’t climb much. Cards that pass the 7-day check are pre-selected for cards the player partially knew; the ones that fail demote back to Wrong and re-enter the funnel at the same low rate. The expected retention curve for hard cards is flat, which is consistent with the desirable-difficulty account: boundary cards need many encounters to consolidate.

Consideration 3: demote-by-one allows a stuck-card edge case

Because a miss demotes a card by exactly one stage rather than retiring it, a card can in principle bounce between Wrong and Known indefinitely:

1. Pass the 7-day check (Wrong → Known)
2. Miss the 90-day check (Known → Wrong, 7-day cooldown)
3. Pass the 7-day check again (Wrong → Known)
4. Miss the 90-day check again (Known → Wrong, repeat)

Call these stuck cards. They never reach mastery, so a topic’s effective active size can creep up over time. I re-read the transition function while updating this article, and there is still no escape valve in it: no counter for Wrong-Known round trips, no forced exit. The clean fix, if the data ever shows stuck cards accumulating, is an explicit valve: after N round trips, force a Mastered transition or a topic-level reset. The loop is a structural property of the state machine, and the same simplicity that allows it makes it easy to detect and fix.

How LearnClash Uses the Retention Curve in Practice and Duel Modes

We hide the retention curve inside gameplay. Solo (the app’s practice mode) is a 6-question session that serves due cards first; Duel mode gives an SRS-due card the same top selection priority as a brand-new one across the 18-question game. The curve runs while you play, which is the difference from flashcard apps where the SRS greets you as a review-backlog chore.

Figure 11: Solo runs the SRS in the open; Duel mode runs it under the surface. The curve drives both.

Mode	Cards per session	SRS visibility	Primary purpose
Solo (topic or category)	6	Foreground, per-question Review pill	Pure SRS exposure, no opponent
Daily Mix	6	Foreground	One-tap sweep of the most-overdue cards, any topic
Duel	18 (6 rounds × 3)	Invisible	Competitive recall + SRS

In Solo, the server composes each 6-question session with a fixed cascade: every due card first, most-overdue leading, then unseen questions in the ladder’s difficulty mix, then cards still on cooldown (soonest-due first), then mastered cards as shuffled filler. Cooldown and mastered re-serves are marked as reviews in the session and don’t advance a card’s stage; answering a cooldown card correctly pays out a bonus Mem instead. Daily Mix runs the same idea across the whole account: it pulls your most-overdue cards from every topic you’ve touched and tops the session up from your recent ones.

In Duel mode, the curve runs invisibly. The selector’s priority cascade treats a card that’s due for review and a card you’ve never seen as equal first-class candidates, prefers questions fresh for both players, and only reaches for on-cooldown or mastered cards when a topic’s supply runs thin (mastered repeats are capped per round). Your opponent can’t tell which of your questions are scheduled reviews, and neither can you. The wrong answer you gave in front of an opponent is the one you remember; the SRS quietly books the follow-up for next week.

There’s no separate review inbox, no due-card counter nagging from a badge. The curve stays underneath the game and decides which cards surface, which is why the 3-stage SRS spaced repetition foundations play out as a game loop instead of a review chore.

Practice your study technique →

The Bottom Line

The 3-stage SRS retention curve is a finite, performance-based version of the classic forgetting curve. Wrong cards return at 7 days, Known cards at 90, and two clean passes retire a card to a sentinel date that will never come due. The checkpoints sit inside the optimal-gap band from the Cepeda spacing dataset, and demote-by-one avoids the ease-hell failure mode SM-2 produces. The cost of that simplicity is gentle punishment for a miss and a theoretical stuck-card loop, both of which I’d rather own than an ease factor nobody can explain.

Figure 12: Four design choices that reinforce each other: the 7-day check, the 90-day check, the sentinel-date exit, and the demote-by-one rule.

If you want to read	Where to go
The broader LearnClash design (ELO, difficulty, modes)	LearnClash statistics page
How ELO-matched wins compound the 7-day SRS check	ELO matchmaking & the balanced win-rate band
The spacing-effect history and Leitner-to-Wozniak lineage	Spaced repetition foundations
The full nine-method study routine using the curve	How to study effectively
How Quizlet’s “SRS” actually works (it isn’t real SRS)	Does Quizlet have spaced repetition?
Why QuizDuel’s 2026 mode roster (Arena, Solo Quests, Events) contains no scheduled-review loop	LearnClash vs QuizDuel

More on the research behind all of this lives in the learning science cluster, starting with the testing effect.