How Often Should a Learner See a Word They Keep Forgetting?

Every cohort has this student. He decides that ubiquitous will not defeat him: he reads it, stars it, drills it — eleven times in one evening. The next morning he opens the app, sees the word, and feels his stomach drop. Nothing. As if they had never met.

Anyone who has crammed vocabulary for an exam knows that feeling, and most learners draw exactly the wrong conclusion from it: "I need to repeat it even more." Massed repetition feels like learning because the word becomes instantly available — but that availability lives in working memory, not long-term memory. Cramming is the most convincing way ever invented to feel like you are learning without actually learning.

We build test-prep apps at Zinuk, an Israeli exam-preparation school. One of them teaches vocabulary for Amirnet (אמיר"ם) — Israel's computerized English proficiency exam for university admission — and its word bank holds 2,725 words, each running on its own private schedule. When we designed that scheduler, the hard question was never which words to teach. It was when: when to show a word for the first time, when to test it, and what to do the moment a learner gets it wrong.

What the science has known for 140 years

Ebbinghaus published the forgetting curve in 1885, and it is still the most important chart in education: memory decays steeply right after learning, and the decay flattens with every successful review. The practical lesson is not "review a lot" — it is review at the right moment: just before the word slips away, not just after you met it.

The second finding is more surprising: the test itself is the learning. Roediger and Karpicke (2006) showed that attempting to retrieve something from memory improves long-term retention more than re-reading the same material — even though re-reading feels easier and safer. Karpicke and Roediger (2008), in Science, pinned down retrieval as the critical ingredient: retrieval practice wins even when it gets less total study time.

What about the exact shape of the intervals? Here the literature is humbler than the marketing around it. Karpicke and Roediger (2010) compared expanding schedules against fixed ones and found a smaller difference than everyone expected. The lesson we took is modest and useful: the big win is spacing plus retrieval. The precise curve is fine-tuning — and fine-tuning should be done on real data, not on ideology.

Two more findings shaped our system. Ellis and Beaton (1993) showed that recognizing a foreign word (L2 → your language) is substantially easier than producing it (your language → L2); they are different skills, and the hard one builds on the easy one. And Kornell (2009) showed that spacing beats cramming even in humble flashcard practice — these effects survive contact with the real world.

What we built — and why it isn't "Anki with nicer colors"

An introduction is not an ambush

A new word is shown once, calmly, with its translation and context — and it is never quizzed immediately afterwards. The reason: immediately after exposure, everyone "remembers". That is the illusion of knowing — confidence peaks at exactly the moment the measurement is worthless. An instant quiz would teach the system that the word is known and teach the learner that he is a genius. Both would be wrong. The first real test comes only after a real break.

Recognition first, recall later

Following Ellis and Beaton, every word climbs a ladder: first recognition questions ("what does ubiquitous mean?"), and only after it has accumulated successful reviews — production questions ("how do you say 'found everywhere' in English?"). A successful production attempt also earns the schedule more credit than a successful recognition: it is stronger evidence of knowledge, so the next interval grows more. Learners experience this as natural gradual difficulty; under the hood it is a ladder of evidence.

Intervals stretch — errors cut

A word answered correctly again and again drifts away: a day, several days, then a multiplier that grows with every success. A forgotten word snaps back hard. And there is one special case we learned to respect: failing a word's first test — a sign that the introduction simply never landed — brings the word back within the same session, not tomorrow. There is no point waiting a day to rediscover what you already know: that word needs one more meeting now.

The "while it's hot" requeue

A word missed mid-session doesn't vanish until the next study day. It re-enters the queue — a few cards back, not immediately. Immediately would be working memory: you would answer correctly and learn nothing. A few cards of distraction are enough to make the second retrieval a real retrieval. At the end of the session, every missed word gets one more closing round. Learners know the feeling as "wait, you again?" — that is not a bug; it is the feature.

The load gate: sometimes the right number of new words is zero

The least popular rule in the system: when the review queue swells past a threshold, new words stop — even if the learner is eager. A postponed review queue is precisely the set of words on the edge of being forgotten; adding new words on top is pouring water into a leaking bucket. Even on a normal day, new words per session are capped at a single-digit number. Progress feels slower; it is faster.

A tuning story: ease hell

We did not invent the scheduling core. Our starting point was the SM-2 algorithm by Piotr Woźniak (SuperMemo) — the workhorse of spaced repetition for thirty years. But anyone who runs textbook SM-2 on exam students quickly meets a phenomenon the Anki community calls ease hell: each item carries an ease factor that drops with every lapse, and in the canonical version its floor is 1.3. A genuinely hard word grinds down to that floor — and then returns again and again, its interval barely growing, forever.

We changed two things — directionally; the exact values stay ours. First, we raised the floor: the canonical 1.3 punishes too hard, and a difficult word in our system keeps its ability to climb out of the pit. Second, we built a rescue mechanism: a word that keeps failing in a row stops accumulating punishment — the system forgives its history and restarts it as if it were brand new. Because at that point, that is exactly what it is.

When is a word "mastered"?

The simple definition — "answered correctly twice" — fails exactly where it matters: it confuses a short streak with durable knowledge. In our system a word is marked mastered only when two conditions hold together: a streak of successes and durability over time — a successful retrieval after an interval of weeks. A word that passed both almost never comes back. A word that passed only one is still on the job.

If you study on your own: four rules

All of these mechanisms — the recognition-to-production ladder, expanding intervals, the while-it's-hot requeue and the load gate — run in production inside Zinuk's Amirnet course, scheduling each of 2,725 words independently. The Hebrew original of this article, with the rest of our learning-science series, lives at The Zinuk Lab.