Forgetting is not the thing going wrong. Memory has two quantities rather than one, and only the first is available to you from the inside: whether a word can be reached at this moment. The second is how well it is actually learned, and the two move separately. A word can be sitting at the front of your mind and barely stored at all, which is what the morning after cramming feels like, and a word can be well stored and completely unreachable, which is what a language you dropped ten years ago feels like on the first day back.
Every practical rule follows from one fact about how those two interact. A review adds most to the second quantity precisely when the first has fallen, so the whole job of a schedule is to reach you shortly before you would have failed, and never earlier.
The six findings this cluster rests on
- testing beats rereading. Trying to recall something is several times more effective than looking at it again, and it feels worse, which is why people avoid it.
- expanding intervals. Review after a day, then three days, then a week, then a month, each time just as it becomes hard.
- desirable difficulty. The effort of retrieval is what does the work, so making study easier usually makes it weaker.
- a word needs about seven encounters spread over time before it is usable.
- interleaving beats blocking. Mixing topics in one session is harder and retains better than finishing one before starting the next.
- sleep consolidates, so a review before bed and a short one in the morning outperform the same minutes at midday.
Two quantities, and only one of them is what you want
Robert Bjork and Elizabeth Bjork set this out in 1992 as storage strength and retrieval strength, and the useful part is not the names. It is that the gain from a successful retrieval is largest when retrieval was hardest, so effort is not a side effect of learning here, it is the mechanism. The reason this is so hard to believe is that it makes the internal signal exactly wrong: the session that felt fluent was the one where least happened.
Two older results sit underneath it. Hermann Ebbinghaus measured his own forgetting in 1885 and found it steep at first and then flattening, which is why the second review has to come within days and the sixth can wait months. Henry Roediger and Jeffrey Karpicke showed in 2006 that being tested on material beats studying it again by a wide margin at a delay, even though the people doing the studying were more confident they had learned it.
What actually happens to one word
The route is not a line and the backward stops are not failures. A word is met inside a sentence you half followed, looked up, and is briefly available, which is the most misleading moment in the whole thing. It is gone by the weekend. It is met again and recognised with effort, and that meeting was worth something because the effort was real. Then it is reached for in speech and missed, which surprises people, and should not: recognising and producing run on separate routes and only the one you practise gets built.
There is a second reason a deck word fails in a conversation, and it is not weakness. Endel Tulving and Donald Thomson showed in 1973 that what gets stored includes the circumstances of storing it, so a word learned in a deck is a word that is retrievable in a deck. That is not a flaw in flashcards. It is an argument about where the last repetition belongs, and the answer is inside a sentence you needed to say.
The other thing worth knowing before you build anything is interference. Learning a set of closely related items together, ten colours, or a word and its opposite on the same evening, makes every one of them harder rather than easier, because what has to be stored is the difference between them and you have supplied nothing but similarity. Split them across weeks. This is the single most common way a well-built deck turns into making cards that teach rather than cards that pile up failing at the first word.
The deck that only ever asks you to recognise
Here is where a year disappears. A deck is built with the foreign word on the front and your own language on the back, reviewed faithfully every day, and at the end of it comprehension has moved a long way and production has not moved at all. The reviews were real. They were just all in the direction you already had.
Reversing the cards is the obvious fix and it fails for a specific reason: a card showing one word in your own language has several correct answers, so you fail it for reasons that have nothing to do with memory and it becomes one of the cards you dread. The version that works is a sentence you have met, with the target word removed. It has one right answer, it tests production, and it carries the context that a bare pair does not. That is the whole difference between a deck that survives a year and one that gets abandoned, and it is the practical content of reviewing a word exactly as you are about to lose it.
One word
From first meeting to arriving without being asked
The route doubles back, and every backward stop on it is where the strengthening actually happened.
- 01Met inside a sentenceHalf understood, and nothing stored yet
- 02Looked up, and availableThe most misleading moment in the whole route
- 03Gone by the weekendThis is the route working, not the route failing
- 04Recognised, with effortThe first review that was worth anything
- 05Reached for in speech, and missedRecognition was built. Production is a separate route
- 06Produced once, slowlyIn a real sentence, to somebody who was waiting
- 07Gone again, more slowlyThe interval grew, which is what progress looks like from inside
- 08Arrives unaskedOnly ever after the doubling back, never instead of it
What the route leaves out is that a few thousand of these are running at once, each at a different stop, which is the whole reason the schedule has to be kept by something that is not you.
Two things worth having on the desk
The Leitner box is the one nobody suggests any more and it is still right for a small collection. Sebastian Leitner described it in 1972: five compartments, a card moves forward when you get it and back to the front when you miss it, and the front compartment is seen daily while the back one is seen rarely. It replaced reading the list again, which is recognition wearing the clothes of study. Free, in the sense that it is a shoebox, and under about two hundred items it beats any application, because the entire backlog is a stack you can see and pick up rather than a number that accuses you.
Above that size the useful feature is not the scheduling, which every program has, but the leech flag. Anki marks a card automatically once you have failed it eight times, and in 2026 that flag is still the most valuable thing in the program. It replaced noticing on your own that one card has beaten you repeatedly, which nobody does, because each individual failure feels like today's problem. Free with the program. When the flag appears, do not try harder.
The mistake: keeping a card that has more than one right answer
A card you have failed eight times is almost never a hard word. It is a badly built card, and the fault is nearly always the same one: more than one answer would be correct, so you are being marked wrong for producing a synonym, or for a card whose front does not narrow the field to a single word. Grinding it is the wrong response and it is the one everybody chooses, because deleting a card feels like giving up on a word.
Delete it or rewrite it as a sentence with a gap. Both are cheap and both end the problem. A handful of these cards can consume a third of every session while teaching nothing at all, and they are also the reason people conclude that spaced repetition does not suit them, when what did not suit them was nine sentences they wrote badly a year ago. The mechanism they are actually fighting is in the curve everyone is fighting, and what bends it.
What this does not settle
Everything above is about a word you have already decided to keep. It says nothing about where the words come from, and a deck built from a frequency list rather than from your own reading produces a schedule you resent, because none of the items arrived attached to anything. The supply side is filling your day with the language without moving country, and the last repetition, the one that moves a word from a card into speech, cannot happen alone at all, which is finding someone to speak to, and what to do in the hour. If the deck is working and the language still is not moving, the cause is further up and is usually one of the named ones in the stalls every learner hits, and the way through each.