Learning design
How MemDecks decides when to show a card again
MemDecks does not schedule repetitions. There is no interval, no ease factor, no next-review date, and no "due today" number anywhere in the product. Instead every card is created with a fixed practice budget — we call it card HP — computed from how hard that specific word is in that specific language, and every correct answer in a game spends some of it. A card is finished when its budget is spent. This post explains the whole mechanism, including the numbers, and says plainly what the design gives up.
The mechanism, in full
There are exactly two numbers on a MemDecks card.
Card HP is the practice budget. It is set once, when the card is created, and never changes. It is computed as round(difficulty tier × 4.2), floored at 100 and capped at 500.
Score is practice credit. Games report a score change after each answer; MemDecks adds it and clamps the result between zero and the card's card HP.
Progress is score ÷ card HP. That is the entire model. There is nothing else — no hidden state, no decay, no timestamps that matter.
Where the difficulty number comes from
The difficulty tier is a number from 0 to 100 computed offline, per word and per study language, before any learner sees the word. It weighs several things: how rare the word is in ordinary use, how long it is and how awkward its spelling is in the target script, how many distinct senses it carries, how morphologically complex it is, and how much it resembles its English counterpart — a cognate is genuinely easier and the score says so.
The resulting scale is grouped into four zones: words a child would know, words a learner meets, words an experienced speaker uses, and words that stay in native-speaker territory. Short frequent words land at the bottom and receive the 100 HP floor. Rare, long, script-heavy words land near the top and receive roughly 400 HP.
Two consequences follow, and both are intentional. The same English word can get different card HP in different study languages, because "water" is not equally hard in Spanish and in Japanese. And difficulty is a property of the word, not of your performance on it — MemDecks does not ask you to rate your own recall, because self-rating is the least reliable input in the whole system.
What one correct answer is worth
Each game decides, and MemDecks adds whatever the game reports. It does not normalise between games, and that is a deliberate refusal rather than an oversight: the games ask for different things, and pretending a lucky multiple-choice pick equals a spelled-out word would make the number lie.
| Game | Score for one answer |
|---|---|
| WordBase Quiz | 3 for a correct first pick, minus 1 per wrong pick in that round, never below 0 |
| Word Type | The number of letters you never had to reveal, reduced if you used the flip or audio hint |
Do the arithmetic and the model becomes concrete. A 100 HP card needs about 34 flawless WordBase Quiz rounds to reach 100%. A 400 HP card needs about four times that. Difficulty is expressed as required repetitions — the harder the word, the more times you must meet it — which is roughly what a learner already believes about difficulty.
Why not just schedule the reviews
Because the failure mode of scheduling is not academic, it is behavioural, and it kills more vocabulary projects than forgetting does.
Interval-based spaced repetition is well evidenced and I am not disputing it. Its problem in practice is that it manufactures a daily obligation, and an obligation converts absence into debt. Miss a week and you return to a queue of overdue cards that is larger than the one you left. For a learner who reviews every day, that is a strength: the system protects them. For a learner whose week got eaten — which is most learners, most of the time — it is a wall, and it arrives at exactly the moment their motivation was already weakest.
MemDecks makes a different bet. If nothing accumulates while you are away, coming back costs nothing, and coming back is the behaviour that actually determines whether anyone learns anything.
What this design gives up
Three real things, stated so nobody has to discover them by disappointment.
- Timing. MemDecks does not place your repetitions where the forgetting curve says they belong. If you practise a card five times in one sitting, MemDecks counts five, and the spacing literature is clear that five spread across two weeks would be worth more.
- Exam efficiency. If you must retain 3,000 items by a fixed date, a scheduler will get you there with fewer total reviews. Use one. The Anki comparison says the same thing in more detail.
- Certainty about recall. A card at 100% means the budgeted practice happened. It is a record, not a prediction, and MemDecks never claims otherwise.
What partly recovers the loss
Two properties of the design do some of the work that scheduling would have done.
First, games draw from your whole collection rather than a review subset, so old cards keep reappearing without anyone scheduling them. Spacing emerges from the size of the collection: the more you own, the longer the natural gap between meeting the same card twice.
Second, the games ask for different things — recognise, spell, match, read in another script — and varying the retrieval route is one of the more robustly useful things a learner can do. A card that reached 60% through several different games is in better shape than one that reached 60% through one, and mixing games is the cheapest improvement available inside MemDecks.
Would this ever change
Possibly, and the shape of a change is already clear: keep the no-backlog property, but let the order in which games draw cards favour ones that have not been seen recently. That would add spacing without adding debt, because a card never becomes overdue — it merely becomes more likely to come up.
Nothing of the kind exists as of 2 September 2026, and this blog does not describe unbuilt features in the present tense. What is live is what is described above.