Applying SM-2 to Chess: Review Tactics, Not the Same Puzzles

For a long time I just solved random puzzles. My rating crept up, but I never actually knew which tactics I was weak at. I knew what a pin was and still kept walking into pins in real games, and no amount of puzzle-solving told me where that blind spot was. I built this site to fix that, and along the way I ran into one problem that can arise when spaced repetition (SM-2) is applied to chess.

The problem with repeating the same puzzle

SM-2 is straightforward to apply when a flashcard asks for a specific answer. The card is the review item, and its answer is the thing to recall.

Chess puzzles are different. Solve the same puzzle again a few days later and it becomes hard to tell whether you recognized the tactic afresh or remembered the answer to that specific board. Remembering "the d5 knight goes to c7 for a fork" is not the same skill as spotting a fork shape on a board you have never seen. Real games require the second skill, so repeating only the same puzzle does not test whether recognition transfers to a new position.

This site's approach: schedule by tactic, not by puzzle

So this site schedules reviews at the level of the tactical concept, not the individual puzzle. Fork, pin, skewer, and every other theme each carry their own interval and easiness factor (EF).

• General training selects due concepts first. When none are due, levels and accuracy help choose new concepts or extra practice.
• Recently attempted puzzles are excluded first, but may repeat when the available pool is too small.
• So what repeats is the pattern itself, not the moves of one board.
• A due concept is evaluated once per review day. Extra practice is recorded without extending its interval. Review days reset at 00:00 UTC (09:00 in Korea).

Left: repeating the same puzzle can mix recognition with memory for its moves. Right: a different puzzle with the same theme tests whether you can find the pattern in a new position.

This structure gives me a practice record for each motif. A low score can suggest what to review, but a review date alone does not identify a weakness in real games. The dashboard radar combines tactic level, accuracy, and EF into six categories. It is not an accuracy-only chart.

A caveat: this does not guarantee improvement. Concept-level scheduling only decides what to review and when. Recognizing the pattern over the board still takes repetition, and how much it helps varies from person to person.

How SM-2 chooses intervals (briefly)

The SM-2 engine underneath is simple. It keeps three numbers per item:

1. Interval: how many days until the next review.
2. Easiness factor (EF): an interval multiplier updated from review results. It starts at 2.5, never falls below 1.3, and can rise above 2.5 after high ratings.
3. Repetitions: successful reviews in a row; it decides when the 1-day and 6-day steps apply.

Each review gets a grade from 0 to 5 and the schedule reacts. In the original SM-2 the learner picks the grade; this site derives it from whether you solved the puzzle, how many hints you used and how long you took. A grade below 3 counts as a failure. The first success sets the interval to 1 day, the second to 6, and after that each interval is the previous one times the EF. An easy item (high EF) spreads out fast (6, 15, 38 days); a stubborn one (low EF) crawls (6, 8, 10). On this site, a failed review resets the interval to 1 day and also lowers the EF, so it comes back more often. (The original SM-2 description leaves the EF unchanged on a failed review.)

Intervals expand after each success: an easy item (EF 2.5) grows 1, 6, 15, 38 days; a hard one (EF 1.3) crawls 1, 6, 8, 10. (Illustrative SM-2 schedule.)

Piotr Wozniak designed the algorithm in 1987 and documented it in full in 1990. It shaped the schedulers that flashcard tools have used ever since. Anki now classes SM-2 as its legacy scheduler and ships FSRS, a retrievability model, alongside it, so "SM-2 runs everything" is no longer accurate; still, many schedulers, Anki's original among them, still take their interval logic from it.

What memory research says about spacing

Spaced study often produces better later recall than massed repetitions, a result reported across many memory studies. Cepeda and colleagues' 2006 meta-analysis focused mainly on verbal-recall tasks and found that the study interval and the final retention interval work together. The analysis does not establish one universally optimal gap or guarantee the same retention in less total time.

This site applies the general idea to chess tactics. The verbal-memory literature does not directly test transfer to over-the-board chess. Showing a different position for the same theme can test recognition without repeating an answer sequence, but the size of any effect on game performance still requires separate data.

References

• Wozniak, P. (1990). SuperMemo SM-2 algorithm, the original specification of the algorithm used here.
• Cepeda, N. J., et al. (2006). Distributed practice in verbal recall tasks: a review and quantitative synthesis, a meta-analysis quantifying the spacing effect.
• Anki Manual: Deck Options, Anki's official distinction between FSRS and its legacy SM-2 scheduler.
• Lichess Puzzle Database, the source of the puzzles scheduled by SM-2 here.