You give a learning tool minutes, not hours. Sponge & Stone runs the most advanced memory model in open research , then goes past it: every review session is composed, card by card, to buy the most lasting memory your minute can afford.
Spaced repetition is one of the most replicated effects in cognitive science. The software mostly stopped improving decades ago. We didn't.
Hermann Ebbinghaus measures how memory decays and how each well-timed recall flattens the fall. Every review app since is a footnote to this experiment.
SuperMemo's SM-2 turns the curve into a scheduling rule. It works, and it never learns: the algorithm most flashcard apps run today predates the web.
FSRS: a memory model fitted to hundreds of millions of real reviews, where even the shape of the forgetting curve is learned from evidence, not assumed.
Running the best published model is table stakes. Here is what we built on top of it.
Every card carries a stability (how long the memory holds) and a difficulty (how stubbornly it resists), updated by the newest generation of FSRS on every answer. Intervals stretch from a day to a season exactly as fast as your memory earns them.
Miss a card five times and we flag it as possibly badly written, then offer to rewrite it, rest it, or retire it.
Stability in days is the interval: the model predicts 90% recall right when the card comes back.
Buttons measure confidence. The clock measures fluency . A solid answer that lands faster than its format's speed bar earns the model's highest rating by itself, and every format gets its own bar, because a map click and a typed sentence take honestly different time.
Fewer buttons, better data. Your answers are timed while you review.
Schedulers stop at "due." Ours keeps going: every due card is priced by predicted retention gained per minute , running the memory model forward both ways, if you recall it and if you don't. A certain success teaches almost nothing. A memory on the verge of being lost is where your minute multiplies, so rescues come first.
The price is divided by your own pace on that kind of question, topics stay varied, and after two stumbles the session deals steadier ground. Every pick can explain itself, in numbers.
Most systems are blind to everything but the quiz. Ours counts proof : hear a fact retold in your private podcast or use it to crack a puzzle, and its schedule renews.
Graded and ungraded evidence stay separate in the ledger, so the retention numbers stay accurate.
The three marks you keep seeing, ○ ◐ ● , are how far a memory has ripened. Every card wears one, so you always know what your minutes have bought.
New and still absorbing. Reviews come quick and close together, because young memories fade fastest.
Setting. Real passes are behind it now, so it comes back less and less often as the memory firms.
Retained for the long haul, and still quietly renewed. If a stone card ever misses, the schedule renewed it too late, and we use that miss to improve it.
Every question is asked with a prediction attached: the recall the model expected, snapshotted at that moment. Then reality answers. Prediction versus outcome, on every single review.
Pass rates by memory maturity, by knowledge shape, by question format. When a slice underperforms, we fix the cards or the model.
The composer shipped against a referee that already existed. If a scheduling idea doesn't show up in retention, it doesn't stay.
Any card in any session can show exactly why it was dealt, with the numbers behind it.
Further reading:
Ebbinghaus (1885), the original forgetting-curve experiments
The open spaced repetition benchmark, where schedulers compete on real review logs
Quantum Country, the essay that puts spaced repetition inside the reading