Spaced Repetition for Chinese, Explained (and Why FSRS Beats Anki's Defaults)

Learning Chinese to a useful level means holding two to five thousand words in memory — words with no cognates, arbitrary-looking written forms, and tones that English ears merrily discard. Nobody's raw memory does that. What makes it tractable is spaced repetition: the one study technique with decades of consistent evidence behind it, and the quiet engine inside every serious Chinese learner's routine.

This guide explains the mechanism from the ground up, and then gets into the part most explanations skip: the difference between the 1980s-era scheduling algorithm most apps still use and FSRS, the modern one — because for a language this size, the algorithm choice is measured in hours of your life per month.

The forgetting curve, briefly

After you learn something, recall probability decays — fast at first, then slower. Ebbinghaus measured this in 1885 and the shape has held up ever since. Two things bend the curve in your favor:

• Retrieval practice. Actively recalling a memory strengthens it far more than re-reading it. A flashcard is a retrieval machine.
• Spacing. Each successful recall after a delay makes the memory decay more slowly. Reviews therefore work best on an expanding schedule: 1 day, 3 days, a week, a month…

A spaced repetition system (SRS) automates this: it tracks every card and asks you each one at (ideally) the moment you're about to forget it. Review too early and you waste time on things you knew; too late and you're relearning from scratch. The scheduling problem — when exactly is "about to forget"? — is the entire game.

Why Chinese is the perfect use case

SRS helps any memorization, but Chinese is unusually well-suited to it, for three structural reasons:

• Every word is a three-way arbitrary binding. Form (学), sound (xué, second tone), meaning ("study") — none predicts the others for a beginner. That's three associations per word to keep alive, which is exactly the kind of load that decays without scheduled retrieval.
• No cognate discounts. An English speaker learning Spanish gets thousands of words nearly free. In Chinese, close to zero — so the memorization volume that SRS optimizes is the dominant cost of the whole language.
• The long tail is genuinely long. Past ~1,000 words, new vocabulary appears too rarely in daily practice to maintain itself through exposure alone. Scheduled review is what keeps word #2,000 alive during the months before you meet it in the wild.

Classic scheduling: SM-2 and Anki's defaults

Most SRS tools — including Anki's traditional scheduler — descend from SM-2, an algorithm published in 1987. It's beautifully simple: each card has an "ease" multiplier (starting at 2.5), and every successful review multiplies the interval by it. Fail, and the card starts over.

Simplicity has costs, and Chinese learners hit all of them:

• Fixed formulas, not your memory. SM-2's parameters were hand-tuned decades ago. It doesn't learn from your review history — a card you've failed five times and a card you've aced five times follow the same arithmetic, just from different points.
• "Ease hell." Each failure permanently lowers a card's ease multiplier. Hard cards — and Chinese has plenty: near-synonyms, similar-looking characters — spiral into punishingly frequent reviews that never relax, even after you've finally learned the word.
• No retention dial. You can't tell SM-2 "I want to remember 90% of my words" and have it schedule accordingly. The interval math is the interval math.

FSRS: scheduling as a memory model

FSRS (Free Spaced Repetition Scheduler), developed by Jarrett Ye and collaborators from research on large-scale review logs, takes a different approach: instead of a formula, it maintains an explicit model of each memory with three quantities:

• Retrievability — the probability you'd recall the card right now;
• Stability — how slowly that probability decays (the interval at which recall drops to 90%);
• Difficulty — how much each review changes stability for this particular card.

Every review updates the model, and the model's parameters are fitted to real review histories — including, as you accumulate reviews, yours. The practical differences:

That last row is the one that matters for Chinese. Reviewing an HSK 4–5 vocabulary of a few thousand cards, a 20–30% efficiency gain is on the order of several hours a month — every month, for years. Anki itself now ships FSRS as an option (off by default for existing users), which is about as strong an endorsement as an incumbent algorithm ever gives its successor.

What this looks like in practice

With FSRS-based scheduling, a well-run Chinese study routine settles into a rhythm: 10–20 new words a day, a review queue that hovers around 50–150 cards depending on your stage, and — the part learners notice most — hard words that stop haunting the queue once you've actually learned them. The scheduler's job is to make the daily cost boring and predictable; your job is to show up.

What spaced repetition can't do