The forgetting curve, and why review beats re-study
Forgetting is the default, not the failure
Every language learner has had the experience: a word you looked up on Monday, used confidently on Tuesday, and could not produce on Friday. The usual explanation people reach for is a personal one — bad memory, not enough effort, too old for this. It is almost never the right explanation.
Retention of newly learned material declines over time unless the material is deliberately revisited, a pattern described by Hermann Ebbinghaus and known since as the forgetting curve. The decline is steep at first and then flattens. Nothing about that is unusual or a sign of a defective memory; it is what memory does with information it has no evidence is worth keeping.
Which reframes the problem. Nobody gets to avoid forgetting, so the only question left is what to do about a curve you cannot switch off.
The two things people do instead, and why they cost more
The first common response is re-study: read the word list again, re-read the grammar explanation, watch the lesson video a second time. It feels productive because it feels easy — recognition is smooth, and smoothness gets mistaken for knowing. But recognizing a word on a page is a much weaker performance than producing it in a conversation, and the ease of re-reading is precisely what makes it a poor guide to whether you have learned anything.
The second common response is volume: more input, more words, more hours. This works in the sense that a larger pile leaves more residue, but it is expensive, and it puts every item on the same schedule — namely, whenever you happen to encounter it again. Items you already know get reviewed as often as items that are slipping away.
The alternative that is cheaper and works better is retrieval: instead of looking at the material again, try to produce it from memory, and only then check. Retrieval practice — the testing effect — consistently outperforms passive review of the same material, and the reason is intuitive once stated: you are practicing the operation you actually need, which is getting something out, not putting it in again.
Spacing: the other half
Retrieval works better when the attempts are spread out rather than bunched together. Spaced study sessions produce better retention than massed ones, which is the phenomenon behind every flashcard scheduler ever built.
The scheduling that follows — intervals that widen while you keep succeeding and collapse the moment you miss, tracked separately for every item — is the subject of homework built from your own mistakes. What that mechanism does not solve is the harder case: what happens when the review has to take place inside a spoken conversation rather than on a card.
Why speaking is the hard case
Flashcards are the easy case, and they are easy for a reason worth naming. The item is presented, you either retrieve it or you do not, the outcome is unambiguous, and the whole event was caused by the scheduler. Prompt, attempt, verdict.
A conversation supplies none of that. There is no prompt — the target structure has to arise from the topic. There is no clean verdict — you may produce the structure correctly, produce it wrongly, produce something adjacent that works, or say a sentence in which the question never came up. And crucially, the scheduler does not control whether the attempt happens at all. You do.
That last gap is where spoken review quietly fails, and it has a name.
Avoidance, and why it looks like success
A learner who is unsure of the present perfect does not usually get it wrong. They route around it. The sentence gets rebuilt mid-utterance into the past simple, or into a shape that needs no perfect aspect at all, and what comes out is correct, fluent and entirely uninformative.
Consider what that does to the record. No error occurred, so no error is logged. The item was due for review, and from the transcript it looks reviewed — the topic came up, the learner spoke, nothing went wrong. But nothing was retrieved either, so nothing was consolidated, and the interval widens on the strength of an attempt that never took place. Repeat that a few times and the schedule has confidently retired an item the learner still cannot produce.
This is the specific limitation of any system that learns about you from a transcript: it sees what you said, not what you declined to say. An error is loud. Avoidance is silent, and silence reads as competence.
There is one usable trace of it, which is why it is worth recording separately from errors: the words a learner audibly hunts for and cannot produce. A hesitation, a circumlocution, an outright request for the word — those surface in a transcript in a way a quietly rebuilt sentence never does, and they mark a gap even though no mistake was made. It is a partial signal rather than a solution, and it is better than treating a clean transcript as a clean bill of health.
Making the situation ask the question
Asking the learner to try harder does not fix avoidance, because avoidance is usually not a decision. The counter is to pick situations that require the structure rather than merely permit it, so that routing around it costs more than attempting it.
This is what a well-chosen role-play buys. Viewing an apartment demands comparatives and modals of obligation, because asking about terms consists of exactly that. Arguing for a raise demands the second conditional and the language of concession, because a case you can only assert is a case you lose. Neither situation lets a learner route around the structure and still get what they came for — which is what turns a conversation into a retrieval prompt wearing ordinary clothes.
Announcing the review would break it. Warned that last week’s mistakes are coming back, a learner stops avoiding and starts attempting — which sounds like an improvement and is the opposite, because the avoidance was the diagnostic. You confirm they can produce the structure on request, which was never in doubt, and lose the evidence that they will not reach for it unprompted.
What a learner can do about their own avoidance
The useful part of all this is that avoidance is detectable from the inside, even though it is nearly invisible from the outside.
Watch for the rebuild. When you start a sentence, feel the structure coming, and quietly restructure to something safer, that moment is the signal — and it is the single most informative thing that happens in a lesson, because it locates a gap that no error log will ever show. Say the risky version instead. A wrong present perfect that gets corrected is worth more than a correct past simple that hid the question.
Notice too that the reflex gets stronger under pressure, which means the structures you avoid in a difficult conversation are exactly the ones least ready for real use. That is uncomfortable and it is the point: the whole reason for practicing under time pressure is to find out which parts of your English survive it.
What this means for how you spend your week
A few practical consequences, none of them exotic.
Short and frequent beats long and rare. Practice spread across the week gets more spacings into the same total time than one long session does, and one long session is mostly one exposure however many minutes it lasts.
Do not re-read your notes as review. Cover them and try to produce the item first; look only after you have tried and failed. The failed attempt is where the work happens, not the part you wasted before looking.
Let the schedule choose, rather than choosing what to review by feel. Feel is systematically biased toward the material you like and are already good at, which is the material with the least to gain.
And say the version you are unsure of. Every practical suggestion above collapses back into that one, because a retrieval you dodged is a retrieval that did not happen, however good the conversation sounded.
The cheap version of this is still worth doing
None of this requires a system. A learner with a notebook, a habit of writing down what they got wrong, and the discipline to test themselves at widening intervals is doing the real thing, and doing it well.
What a system adds is bookkeeping people are bad at: remembering that a particular item is due today, noticing that a specific error has now recurred three sessions running, and — the part a notebook genuinely cannot do — arranging for the next conversation to land on a situation where that item has to be produced out loud. That is the mechanical half of the closed learning loop.
The curve is not negotiable. The schedule you run against it is.