The Use and Abuse of Spaced Repetition

By the developer of SiteDictation  ·  2026

Spaced repetition is one of the few things in language learning that is genuinely backed by cognitive science. The forgetting curve is real. Reviewing material at expanding intervals does work. Anki has helped millions of people memorize vocabulary they would otherwise have forgotten. This essay is not an attack on spaced repetition.

It is an argument that spaced repetition, applied dogmatically at scale, becomes the thing holding you back. Not a tool. A trap.

The Sisyphus Problem

Early on, spaced repetition feels like a superpower. You add sentences. You review them. The queue stays manageable. You see progress. This works.

Then, somewhere around 5,000 positions, something shifts. The queue never empties. You do 500 reviews. Tomorrow there are 480. The day after, 510. You work harder — 600, 700, 800 reviews a day — but the mountain does not shrink. You are Sisyphus. The rock is your review queue.

Most learners at this point conclude: I need more immersion. More input. More time living with the language. They are wrong about the diagnosis.

"Lack of real immersion was not really what was holding me back — but actually my own suboptimal algorithm, which was dogmatic spaced repetition, useful up until a point."

I discovered this after about a year of intensive Romanian practice. The problem was not immersion. I had time to practice every day. The problem was that my algorithm spent all of that time on review, and none of it on new material. The queue was so large it crowded out everything else.

What "Dogmatic" Means

Dogmatic spaced repetition has a simple rule: always prioritize review of due positions. Only introduce new material when the review queue is empty.

With 1,000 positions, the queue empties regularly. The rule works. With 10,000 positions, the queue is never empty. There are always some due positions. The algorithm starves you of new experience indefinitely. You drill the same territory endlessly, while whole continents of the language go unexplored.

Traditional systems — SM-2, Anki in its default configuration — follow this rule faithfully. They were designed for flashcard memorization at modest scale. Nobody asked: what happens when the deck has 26,000 cards?

Speaking Power vs. Debt Payment

There are two fundamentally different activities that look identical from the outside: sitting down and doing SRS.

The first is building speaking power — exposure to new sentence patterns, new vocabulary in context, new constructions you have never seen before. This is what makes you better.

The second is debt payment — reviewing positions you already largely know, servicing a schedule because the algorithm demands it, doing extra homework to please a system rather than to improve a skill. This feels productive. It is mostly not.

At scale, dogmatic SRS collapses this distinction entirely. Everything becomes debt payment. The algorithm has become your master, and you are working for it, instead of it working for you.

The Fiat Factor

The solution I arrived at is what I call a fiat factor.

The name comes from monetary economics. When a government runs a deficit, it does not wait until all existing debt is paid before spending on new infrastructure. It coins new money — by fiat — and spends it into existence despite the outstanding obligations. The debt is real, but so is the need to build something new. You cannot wait for the books to balance before the country moves forward.

In language learning terms: at each session, before consulting the spaced repetition scheduler at all, the algorithm flips a coin. On roughly 20–25% of selections, it bypasses the review queue entirely and forces a new, previously unseen position — regardless of how many reviews are pending. On high-immersion days, this percentage rises to 80%.

This is heresy to traditional SRS. The algorithm does not even ask the scheduler for a recommendation. It overrides it.

The result: you always encounter new material, at every stage of learning, no matter how large your review backlog grows. The debt is acknowledged but not tyrannical. You keep building.

LIFO, FIFO, and the Stateless Session

Within the review portion of the algorithm, another insight applies.

Traditional SRS, when you fail a card, makes you retry it within the same session until you get it right. It forces you to clear your failures before you can proceed or stop. This is coercive. It turns 25 minutes of practice into 45, and it ends in resentment.

Instead, the algorithm alternates between two selection strategies: LIFO (most recent failure first) and FIFO (oldest due position first). LIFO handles about 25% of review selections. The effect: if you failed a position two exercises ago, there is a natural chance you will see it again in the same session — while the sound is still in your ear, while the muscle memory is warm. But you are not forced to clear it. The algorithm creates the opportunity probabilistically, without coercion.

This keeps the session stateless. The algorithm does not need to track which positions you failed this session and enforce a retry loop. The LIFO distribution handles it naturally. And the session ends in 25 minutes, cleanly, whether or not every failure was resolved.

What the Data Shows

After four years of practice and over 100,000 recorded transactions, the streak-based intervals are validated by actual performance data:

At four consecutive successes — 96% correct — a position is genuinely mastered. Language knowledge does not decay in six months. Spending review time on these positions is pure debt payment with near-zero return. The algorithm correctly ignores them and redirects that time toward positions that need it.

The 5-minute retry for fresh failures is validated by the recovery data: after a first failure, success rate on the immediate retry is around 72%. The audio is still in working memory. The motor pattern is still warm. Retrying now works in a way that retrying tomorrow does not.

Why Anki Doesn't Do This

SM-2 was designed in 1987 for flashcard memorization. The use case was vocabulary — discrete items, quality ratings from 0 to 5, an easiness factor that adapts per card. It was not designed for native-speaker audio transcription, for corpora of 26,000 sentences, or for the question of what happens when you simply cannot drain your review queue in any reasonable timeframe.

Anki implements SM-2 faithfully, and for what it does, it does it well. The problem is not Anki's faithfulness. The problem is that the original question — what does learning at scale actually require? — was never asked.

A system designed to help you memorize 500 vocabulary words behaves very differently when applied to 10,000 sentences practiced over four years. The algorithm does not know it has become Sisyphean. It keeps scheduling. You keep reviewing. The queue never empties. And the year passes.

The Result

(The terms used throughout — position, transaction, streak — come from portfolio management. If they are unfamiliar, The Ledger Model explains the vocabulary and the architectural choice behind it.)

Using this approach — fiat factor, LIFO/FIFO alternation, 25-minute sessions — I reached C1 in Romanian according to the Common European Framework. That is the level at which you understand demanding, extended texts and can express yourself fluently and spontaneously without obvious searching for words.

I did not achieve this by increasing immersion. I did not move to Romania. I did not find a language partner. I fixed the algorithm.

The method is the app. The app is available for Romanian, free, for macOS. If you are learning a language and you have been practicing for more than a year without feeling like you are actually getting anywhere — it may not be you. It may be the queue.