Spaced repetition, explained properly

In 1885 Hermann Ebbinghaus memorised lists of nonsense syllables, tested himself at intervals, and plotted how fast he forgot them. The resulting curve is steep: most of what you learn in a session is gone within a day, and the majority of the loss happens in the first few hours.

He then noticed something more useful than the decay itself. Each time he relearned a list, it decayed more slowly. The curve flattened. And the flattening was largest when he waited until he had almost — but not quite — forgotten.

That is the whole idea. Everything since has been refinement.

Retention
100% ┤●
     │ ╲                    ← one study session
     │  ╲___
 50% ┤      ╲_____
     │            ╲_________
     └──────────────────────────── time
      
100% ┤●    ●     ●        ●          ← reviewed at the right moments
     │ ╲    ╲     ╲        ╲___
 50% ┤  ╲    ╲__   ╲___        ╲______
     │   ╲      ╲      ╲____
     └──────────────────────────── time
        1d   3d    1w      3w

Each review resets the curve and makes the next descent shallower. Four well-timed reviews produce a memory that survives months. Four reviews in one evening produce a memory that survives until Thursday.

Why the timing matters

The intuitive approach is to review often. Review it every day and surely it sticks harder.

It does not, for a reason that is initially counterintuitive: the difficulty of the recall is what strengthens the memory. If you review a word ten minutes after learning it, retrieval is effortless, and effortless retrieval teaches your brain almost nothing. If you review it after three days, when it is genuinely on the edge of being gone, the retrieval is hard — and the successful effort is what consolidates it.

Cognitive scientists call this desirable difficulty. It has an unfortunate side effect: the more effective method feels worse. Spaced review feels like you are constantly half-forgetting things, because you are. Massed review feels fluent and productive, and produces knowledge that evaporates.

Learners consistently rate the less effective method as more effective. This is worth knowing about yourself in advance.

Retrieval, not exposure

A closely related point that is even more often missed: re-reading is not review.

Reading a vocabulary list again is exposure. Looking at the word, feeling the flicker of recognition, and moving on produces almost nothing. What produces memory is retrieval — pulling the item out of your head, from a cue, with a real chance of failure.

This is why flashcards work at all. Not because they are cards, but because they force a moment where you must produce the answer before seeing it. Any format with that property works. Any format without it — highlighting, re-reading, listening passively to a list — mostly does not.

Which means the single most important rule of using flashcards is: answer before you flip. Say it out loud. If you flip the card while thinking “yes, I know this one”, you got the exposure and skipped the mechanism.

The intervals

A typical schedule for a card you keep getting right:

Review	Interval since last
1	1 day
2	3 days
3	1 week
4	2 weeks
5	1 month
6	2.5 months
7	6 months

Get it wrong and the interval collapses back to a day or two, then re-expands.

You do not need to manage this. Software — Anki and everything descended from it — computes the next interval per card based on your history with that specific card, using a scheduler like SM-2 or the newer FSRS. The exact algorithm is a modest optimisation over the principle; the principle is 95% of the benefit.

The consequence that matters practically: your daily workload falls over time. A deck of 2,000 mature cards does not mean 2,000 reviews. It means perhaps forty, because most of the cards are on multi-month intervals. This is the property that makes maintaining a large vocabulary tractable at all.

Four mistakes that ruin it

1. Isolated words

By far the most common. A card with hablar → to speak teaches you a translation pair. It does not teach you that hablar con takes con, that it is pronounced with the stress on the second syllable, or how it behaves in a real sentence.

Worse, it trains the wrong retrieval path. You end up able to translate a list and unable to produce speech, which is the understand-but-cannot-speak problem in miniature.

Use sentences. One card, one sentence, containing one new item. Everything else in the sentence should already be known.

2. Cards you did not make

Downloaded decks fail at a remarkable rate. Partly because the cards are not about anything you needed, and partly because the act of making a card is itself a learning event — you had to decide what the item was, what the boundaries of its meaning were, and how to cue it.

There is an old Anki maxim: don’t use other people’s decks. It overstates the case for things like the 2,000 most common characters, where the material is genuinely universal. For everything else it is correct.

The best cards come from language you actually needed and could not produce — the sentence you reached for in a lesson and failed. That sentence is already motivated, already contextualised, and already emotionally tagged.

3. Skipping days until the backlog wins

Miss a week and 300 cards come due. The deck now looks like a punishment, so you miss another week, and now it is 600.

Two fixes. First, keep the daily count small enough that a missed day is recoverable — twenty new cards a day is a lot over a year (7,000 cards) and most people should be doing five to ten. Second, when you do fall behind, do not try to clear it. Set a daily cap and let it drain over two weeks, or delete the backlog outright. A deck you dread is worth less than a smaller deck you open.

4. Recognition instead of recall

If your card shows the target-language word and asks for the translation, you are testing recognition, which is the easy direction and the less useful one.

Reverse it. Show the meaning or the context, and produce the target language. Production is harder, which is the point. In practice a mixed deck — some recognition, some production, some audio-only — covers more of the skill than any single direction.

Where it does not help

Spaced repetition is a memory tool. It is excellent for vocabulary, characters, gendered nouns, irregular forms, and set phrases — anything that is genuinely arbitrary and must simply be retained.

It is poor at:

• Fluency, which is speed of access under load, not availability
• Pronunciation, which is a motor skill and needs production practice — see shadowing and minimal pairs
• Listening, which needs volume of real input
• Grammar intuition, which needs exposure to many varied instances, not repetition of a rule

A learner whose entire practice is flashcards ends up with a large passive vocabulary and no speech. Flashcards are the memory layer underneath the actual work; they are not the work.

A realistic setup

For an adult learner with twenty minutes a day:

1. Five to ten new cards a day. Not twenty. This is 2,000–3,500 items a year, which is a functional vocabulary.
2. Sentences, not words. One unknown per sentence.
3. Audio on every card, if you can. It makes the card testable as listening, and it stops you learning a wrong pronunciation for a word you have only read.
4. Cards built from your own lessons or reading, not a downloaded list.
5. Reviews before new cards. If you only have ten minutes, do the reviews and skip the new material. Reviews are the compounding part.

The friction in all of this is card creation — finding the sentence, the translation, the audio, the transcription. It is why decks die in week three. Automating that step is exactly what Teachee does with the sentences from your lessons: they arrive already transcribed, translated and phonetically annotated, and export straight to Anki, so the deck builds itself out of language you actually needed.