The finding that started it
In the 1880s Hermann Ebbinghaus memorised lists of nonsense syllables, waited, and measured how much he had lost. He was his own subject, the material was meaningless by design, and the result was the forgetting curve: retention falls off steeply at first, then flattens. Most of what you forget, you forget quickly.
The second finding is the one that matters, and it is the reason any of this works: every time you successfully retrieve a memory, the curve gets flatter. The fact decays more slowly than it did before. Review it again, later, and it flattens further. Forgetting is not a leak to be topped up; it is a process you can slow down, by retrieving.
The counterintuitive part: difficulty is the point
The instinct is to review often, while the material is fresh, because that feels safe. It is close to useless. If you can already recall something effortlessly, retrieving it again teaches your brain almost nothing — no effort, no signal, no strengthening.
The strengthening happens in proportion to the *difficulty* of the retrieval. So the optimal moment to review a card is as late as possible while still succeeding: right at the edge of forgetting, when it takes real effort to drag the answer up. Reviewing earlier than that is comfortable, feels productive, and is largely wasted time. This is the idea researchers call a desirable difficulty, and almost every intuition you have about studying runs against it.
So when, exactly?
That is the entire technical problem, and it is why algorithms exist. The edge of forgetting is different for every card and every person: you'll hold "the mitochondrion produces ATP" for months and lose an obscure enzyme's cofactor in four days. A schedule that treats those the same is wrong for both.
A real scheduler estimates, for each card, how likely you are to recall it right now, and shows it to you when that probability has decayed to a target — around 90%, typically. To do that it needs a model of your memory of that specific card, updated every time you review it. That is what SM-2 attempted in the 1980s and what FSRS does considerably better today.
How to tell a real one from a fake one
- Ask what happens when you rate a card 'easy'. In a real scheduler the next interval jumps substantially — weeks, then months. In a fake one it appears again on Thursday because Thursday is when the ladder says so.
- Ask whether two cards you rated differently diverge. If a card you find trivial and a card you keep failing are on the same schedule, there is no memory model. There is a loop.
- Ask the vendor to name the algorithm. "Spaced repetition" is not an algorithm. "FSRS-4.5" and "SM-2" are. A product that cannot name one does not have one.
The part nobody mentions
Spaced repetition only works on cards that exist, and writing good cards is slow, unglamorous work. This is the actual reason most people abandon it: not the reviews — the authoring. They buy into the theory, hand-write sixty cards over a fortnight, realise the semester needs eight hundred, and quietly stop.
There are two honest ways out. Use somebody else's deck, and accept that it is somebody else's course, with somebody else's errors, drilled into you by an algorithm built to make things permanent. Or generate the cards from your own material, which is the problem Thaxas exists to solve — and then let real FSRS schedule them.