Spaced Repetition Schedules Explained: Fixed Steps, SM-2 and FSRS

Everyone who starts with spaced repetition eventually asks the same question: what are the right intervals? One day, three days, a week? Is 1-2-4-8 better than 1-3-7-21? Should an algorithm decide?

There is a real answer, and it is less dramatic than the debates suggest. The principle - growing gaps, reset on failure - does nearly all of the work. The exact numbers are a refinement. This guide explains the three families of schedules in use today, what each one actually does with a German word, and how to choose.

The idea all of them share

After you learn something, the chance that you can recall it falls over time, quickly at first and then more slowly. Each time you successfully pull it out of memory, the fall becomes slower. So the gap before the next review can be longer than the last one.

Two more findings shape every schedule:

• Retrieval has to cost a little. A review that is trivially easy strengthens the memory very little. The most productive moment is when recall takes some effort and still succeeds.
• Forgetting is information. A failed review means the gap was too long for this item, and the next one must be shorter.

Every schedule is a rule for choosing the next gap. They differ in how much they know about you and the item when they choose.

Family 1: fixed steps

The oldest and simplest approach. Every item climbs the same ladder of intervals. A correct answer moves it up one step; a wrong answer moves it down.

The paper version is the Leitner box from the 1970s: several compartments, reviewed every day, every other day, weekly, and so on. It is described in the Leitner system for German. The best-known digital version is WaniKani's, which merke.me uses for German:

After the eighth review the item is finished and does not come back.

Strengths. It is predictable: you can calculate your workload for next month. It is transparent: you always know where an item stands and why. There is nothing to configure and nothing to misconfigure. And it has an end, which matters psychologically more than efficiency charts admit.

Limits. It treats der Computer and die Voraussetzung the same on the way up, although one is far easier for an English speaker. Easy items get a few more reviews than they strictly need. The correction happens through failures only: hard items fall back and repeat the short steps.

Family 2: SM-2 and its relatives

In 1987 Piotr Wozniak published an algorithm, SM-2, for his program SuperMemo. Anki used a version of it for most of its history, and many apps still do.

SM-2 gives each card its own multiplier, called ease, starting at 2.5. After each successful review the interval is multiplied by the ease: 1 day, 6 days, 15 days, 38 days, 94 days. How you rate the card adjusts the ease: "easy" raises it, "hard" lowers it, a failure lowers it and resets the interval.

Strengths. Cards drift apart according to their difficulty. An easy card races ahead; a hard one stays on shorter gaps. The intervals have no ceiling, so a card can go years between reviews.

Limits. The well-known one is "ease hell". Every failure lowers a card's ease, and little raises it again, so a card that had a bad month can end up with tiny multipliers and come back far too often for years. The other limit is that the same formula is used for everyone, whatever their memory or material.

Family 3: FSRS and memory models

FSRS, built into Anki since late 2023, takes a different route. It keeps a small model of your memory for each card: how stable the memory is (how long until recall probability drops to 90 percent), how difficult the card is, and how likely you are to recall it right now. It fits its parameters to your own review history.

You then set one number, the desired retention, for example 0.90. FSRS schedules each card for the moment it predicts your chance of recall falls to that level.

Strengths. In comparisons on large sets of real review data it needs clearly fewer reviews than SM-2 for the same retention. It handles late reviews sensibly. And the control it gives you is the right one: you choose how much forgetting you accept, and the workload follows.

Limits. It needs honest, consistent button presses and a few hundred reviews before it is tuned to you. The intervals are no longer something you can predict or explain; they are outputs of a model. And because items never finish, the deck has no end. Settings for German are in Anki settings for learning German.

How much does the choice matter?

Less than the argument around it.

An adaptive algorithm can save a meaningful share of reviews compared with a fixed ladder, and for someone holding 20,000 cards across several subjects that is a real saving. For someone learning 3,000 German words in a year at 10 a day, it is a few minutes a day. The same learner loses far more to cards without articles, or to three skipped days a week.

The old question: 1-3-7-21?

Hand-made schedules like 1-3-7-21-60 or 1-2-4-8-16 are fixed ladders with different rungs. Any ladder that roughly doubles or triples each step, starts within a day, and reaches at least a month will work. Two details matter more than the exact numbers:

• Start early. A first review within hours catches the steepest part of the forgetting curve. Schedules that start at "tomorrow" lose more new words on day one.
• Go long. A ladder that stops at two or three weeks keeps you reviewing words forever, or lets them fade. The step to several months is what makes a word permanent.

What a wrong answer should do

Schedules differ more in how they punish than in how they reward.

• Full reset (classic Leitner, old Anki defaults): the item goes back to the start. Harsh on a word you knew for three months and slipped on once.
• Partial drop (WaniKani, merke.me): the item falls one step in the early stages and two in the later ones. Progress is dented and not erased.
• Model update (FSRS): the stability is reduced by an amount that depends on the card's history.

For German, a partial drop or a model update fits better than a full reset, because many failures are small: you knew the word and missed the article. That deserves a setback, not a restart.

merke.me adds one thing to the partial drop. When you miss an entry that had already climbed a few steps, an extra "ghost" review of it starts at step one and has to be walked up through the short intervals again, while the entry itself only drops. The stubborn word gets dense repetition; its long-term progress is kept. More in why you keep forgetting the same German words.

Which should you use?

• You like to measure, tune and optimise, or you hold very large decks: FSRS in Anki.
• You want to know what tomorrow looks like and never think about settings: a fixed ladder.
• You study from paper: Leitner.
• You are choosing an app: check that it has growing intervals that reach months and that a wrong answer has a consequence. Several popular apps fail that test; see the best spaced repetition apps for German.

merke.me uses the fixed eight-step ladder on purpose. We chose predictability and a finish line: every entry is mastered after eight correct reviews, you can see its step on every card, and the daily workload follows directly from how many new entries you start. The German itself - 10,000 frequency-ordered words, collocations, about 480 grammar points - sits on that one schedule in one queue. If that is the kind of system you want, the decks show what is in it, and 100 words and 30 grammar points are free to learn.

Whichever family you choose, the method only works on the days you open it. How to use it well, whatever the algorithm, is in how to use spaced repetition for German.