How reviews are scheduled

Two questions decide every review: when a card is due, and which due card you see next. FSRS answers the first and a draw answers the second. Every number on this page comes from Lymi’s code when the site is built, so it stays current.

FSRS decides when a card is due

Lymi uses FSRS version 6 with its default weights. FSRS tracks three numbers for each way a card is asked.

Retrievability

The chance you’d recall the card now. It starts at 100% after a review and falls from there.

Stability

Days until retrievability falls to 90%. It grows each time you remember the card.

Difficulty

How hard the card is for you, from 1 to 10. Forgot and Hard raise it, Easy lowers it. The higher it is, the slower stability grows.

A card is due when its retrievability falls to 90%. Stability is measured at 90% too, so a card’s interval equals its stability.

A higher target means shorter intervals and more reviews. The Anki manual warns that the workload climbs fast above 90%. A lower target means fewer reviews and more forgotten cards.

The first grade sets the pace

Each row is a new card: one grade, then Good every time it comes due.

The minutes are FSRS’s learning step. Lymi doesn’t wait for them: a card still learning comes back after a few other cards in the same review.

On a learned card, only Forgot goes back

A card at each Good interval above, graded on its due day. Hard still moves it forward, just less than Good.

Forgetting shrinks stability, not history

Forget a card with an interval of 5 months and it restarts at 5 days, not the 2 days of a new card. FSRS keeps part of its stability, so it climbs back faster.

Real intervals vary a little

FSRS shifts longer intervals by a few days at random, so cards learned together don’t all come due together. The tables show intervals before the shift. For example, 11 days becomes 9 to 13 days, 46 days becomes 42 to 50 days, 5 months becomes 153 to 173 days.

What can come up today

A card can be asked in several review modes, such as term → meaning. Each mode has its own schedule. A mode can come up today if:

• It’s due before your day ends. Your day follows the review timezone in Settings.
• Its card and deck are active, the card uses that mode, and you’re a member of the deck.
• It’s the card’s next mode. Modes start one at a time, once the previous one reaches Review. A mode missing its cue, like a picture mode without a picture, is skipped.
• You haven’t reviewed another mode of this card today. A revealed answer stays fresh for hours, so a second mode would be too easy.
• It hasn’t been forgotten twice today.

The review, Today’s count, deck counts, the end of the day and reminders all use this rule, so they always agree. Modes start with meaning → term, the harder direction. The sources explain why.

What comes next

Lymi stores nothing about a review. After each grade, it picks the next card from your cards, today’s grades and the date. It takes the first match below.

A forgotten card comes back once, after a while

A card you mark Forgot returns once, after about 20 other cards. That’s counted in cards, not minutes, so a break doesn’t bring it back sooner. A return counts toward your daily goal. Hard is a recall, so a card you mark Hard waits for tomorrow like any other.

A second try a minute after the answer only tests what was on screen. Once each in several sittings does more for memory than several tries in one, so the return waits, and a card forgotten again waits for tomorrow. Easy on a card you’ve already seen today counts as Good: the schedule shouldn’t stretch because you saw the answer a few minutes ago.

A return is never shown before its gap. If the review runs out first, the card waits in Review forgotten at the end, a second look you choose to take. A card forgotten twice in a day, or on 3 of the last 5 days it came up, offers to add a hook.

Cards left learning come back first

If you stop while a missed card is still waiting, it stays in learning. The next day the draw shows it first, then one more every 3 attempts, so yesterday’s hard cards don’t pile up at the start.

Ordinary draws mix reviews and new cards

Every fifth ordinary draw is a new card. When reviews run out, every draw is. Returns and cards left learning don’t count toward the 5.

Reviews are weighted by retrievability to the power of 4. Cards you’ll likely remember come up more often, but weak cards still appear. The simulation shows the trade-off.

Of every 4 new-card slots, 3 favour recent cards: odds halve every 7 days after a card is added. The fourth takes your oldest unstarted card, so a big deck you joined keeps moving while this week’s lesson comes first.

Every device draws the same order

The order looks random but isn’t. Each card’s place comes from a hash of the date, the card and the mode. The same cards and grades give the same next card after a reload, offline, on another device or in a single deck. Grades from another device can change what’s next. A new order starts at local midnight.

A day, attempt by attempt

This learner is simulated, not real. They added 120 cards every 7 days and did 50 attempts a day for 45 days. Here is their next day, run through Lymi’s draw and scheduler. Each tile is one attempt, and a card keeps its number all day. Hover or tap a tile to follow that card.

A review stopped partway

The same learner stops after 20 attempts. 3 missed cards are still waiting.

Reopening the review later that day shows card 21, the same card the uninterrupted day showed next. Only the grades were saved. If they wait until tomorrow, the waiting cards come first, 3 attempts apart.

Why these weights

The weights come from a simulated year under each rule Lymi considered. Every rule uses the same draw with a different order, and every grade goes through the real scheduler.

Review order

One learner, 365 days, 50 attempts a day, one new card in 5, averaged over 4 runs. Remembered is the total retrievability of started cards on the last day. Late is cards more than a week overdue.

Highest recall first remembers the most, 6% more than Lymi. But it leaves 41% more cards over a week late and shows strong cards in the same order every day. Lymi leaves the fewest cards late. Lowest recall first, which Anki suggests for a large backlog, leaves the most late when the daily limit never lifts.

Returns

The same learner and year, and 200 short reviews of one 30-card deck played to its end. Each rule swaps only when a forgotten card comes back. A sitting’s returns and gap are averages per review; waiting is the cards left for Review forgotten.

The old rule spent 10.8 of 50 attempts a day on returns, about 6.4 cards after the miss, and in a short review it served them early. One return after 20 spends 6, and no card is seen three times in a day. The simulated learner recalls a return 75% of the time whatever the gap, so this table measures what returns cost, not what a longer gap teaches. That part comes from the research below.

New cards

A 300-card deck joined on day one, a 60-card lesson every 7 days, and 25 attempts a day for 365 days. Cards arrive faster than they can start.

Halving alone starts lessons fast but stalls the joined deck at 118 cards. A floor lets old cards crowd out lessons. The oldest-card slot starts lessons in 6 days at the median and starts 251 joined cards.

What the sources support

No study tests Lymi’s rules as a whole. Each source backs a part, and two point the other way.

• FSRS: The Algorithm, and ABC of FSRS

  Open Spaced Repetition wiki, and ts-fsrs, the implementation Lymi runs

  Supports

  What stability, difficulty and retrievability mean, the FSRS-6 curve, 90% as the default target, and that higher targets mean more reviews.

  Does not

  Which due card to show first, returns, or one mode per day.

• Deck Options

  The Anki Manual

  Supports

  Workload rising fast above 90%, short learning steps like 10 minutes, fixed orders making answers guessable, and hiding a card’s siblings until tomorrow.

  Does not

  Lymi’s review weighting. For a large backlog it recommends weakest cards first, and it gives no evidence for its warning about order.

• Improving sort orders

  Anki Forums, 2024, on Jarrett Ye’s sort-order simulations

  Supports

  Under a daily cap, order matters, and highest recall first spends the least time per card remembered.

  Does not

  A weighted random order, which it didn’t test. It used 80% retention and a heavy backlog, and lowest recall first remembered the most cards.

• Enhancing human learning via spaced repetition optimization

  Tabibian, Upadhyay, De, Zarezade, Schölkopf and Gomez-Rodriguez, PNAS, 2019

  Supports

  Randomized review timing, set by a memory model.

  Does not

  Favouring cards you’ll likely recall. It reviews an item more the likelier it is forgotten, has no daily cap, and uses Duolingo logs rather than an experiment.

• Priority queue

  SuperMemo Help

  Supports

  Strict priority order has flaws, and some randomness helps.

  Does not

  How much randomness to use. Too much undoes the priorities, and its priority is set by the learner, not predicted recall.

• The effects of receptive and productive learning of word pairs on vocabulary knowledge

  Webb, RELC Journal, 2009

  Supports

  Starting with meaning → term. Productive learning beat receptive learning on most tests.

  Does not

  Mode order in spaced review. It studied single sessions, and receptive learning was better for recognising meaning.

Two rules have no study behind them: one mode per card per day, and a random order so position can’t become a cue.

• Decks and cardsHow a card is shaped and which modes it uses.
• API referenceThe review queue and what the draw reads.