The science behind GraphChinese
GraphChinese is Math Academy's approach to learning, built for Mandarin. The learning science it runs on is not new or contested — active recall, spaced repetition, mastery gating, and cognitive-load management have been reproduced in study after study for the better part of a century. What has been missing is a Chinese course engineered around them end to end. This page explains the science and points at the exact mechanism in GraphChinese that implements each piece.
The belief underneath all of it: most people are capable of far more than society expects. When someone "fails" to learn Chinese, the limiting factor is almost never ability — it is an environment that failed to ensure learning. Mastering Chinese is hard, but it is just work, and the work can be mapped.
The bottleneck is memory, so the engine is built around memory
Two facts about human memory shape everything in GraphChinese.
Working memory is tiny. You can hold roughly four chunks of new information at once, for around twenty seconds, and less while you're manipulating them. If a lesson asks you to juggle new characters, new grammar, and half-remembered old vocabulary at the same time, it has already failed — the load exceeds the hardware.
Long-term memory decays on a curve. Ebbinghaus measured this in the 1880s: without review, recall falls off fast, and each successful, well-timed retrieval flattens the curve. Memory is not a bucket you fill once; it is a quantity you maintain.
GraphChinese turns both facts into machinery. All 743 topics — from the four tones and the first stroke of 你 up through HSK 6 grammar — live in one dependency graph, and a topic stays locked until every prerequisite is proven mastered. That gate is not a difficulty setting; it is cognitive-load management. When prerequisites live in long-term memory, they cost working memory nothing, and the whole four-chunk budget is available for the one genuinely new thing in the lesson.
Decay is modeled explicitly. Every topic you've learned carries a memory value that decays as memory · 0.5^(days since review / interval), and a review surfaces the moment that value is about to slip below half — just before forgetting, never crammed the night before. Intervals stretch from one day toward 120 days as your retention is proven, so mature knowledge gets cheap to maintain.
Eleven principles, eleven mechanisms
These are the reproduced findings GraphChinese is built on — the same list Math Academy engineered into mathematics — each with the concrete feature that implements it for Chinese.
1. Active learning
Students learn more from doing exercises than from consuming content.
In GraphChinese: every card is a retrieval act. Lessons are the shortest instruction that enables solving — about 120 words per concept — and then you solve. There is no "watch this video" as a primary learning mode.
2. Deliberate practice
Effective practice is individualized work at the edge of your current ability, chosen to improve specific weaknesses.
In GraphChinese: the task selector only serves topics whose prerequisites you've mastered — your personal frontier — and follows up on weak spots. There is no fixed syllabus everyone marches through.
3. Mastery learning
Students should demonstrate proficiency on prerequisites before moving to advanced material. Bloom's research on mastery approaches in the 1980s found effects large enough that he framed the field's challenge as reproducing one-on-one tutoring at scale.
In GraphChinese: the hard rule. A topic unlocks only when every prerequisite shows a proven mastery score and a completed lesson. No "unlock all" — though you can test out of anything you already know via placement, so mastery gating never means re-grinding.
4. Minimized cognitive load
New information must arrive in pieces small enough for working memory (Sweller's cognitive load theory).
In GraphChinese: each worked example introduces exactly one new element beyond the previous one. Exercise sentences carry at most about two unfamiliar items; everything else is prerequisite material your long-term memory supplies for free.
5. Automaticity
Low-level skills practiced to the point of running without conscious effort stop consuming working memory.
In GraphChinese: rapid-recognition drills — short frames answerable in seconds — build automaticity on tones, characters, and high-frequency words, so grammar and meaning get the mental headroom.
6. Layering
Learning is connecting new material to what you already know; knowledge deepens fastest when new topics exercise recent ones.
In GraphChinese: the dependency graph is layering made visible. Mastering a topic propagates partial credit up to everything built on it, and new topics are chosen because they exercise recently learned prerequisites.
7. Non-interference
Similar material learned back-to-back interferes with recall; new concepts stick better next to dissimilar material.
In GraphChinese: the selector avoids stacking easily confused items — near-synonyms, similar measure words, 的/得/地 — in the same session.
8. Spaced repetition
Reviews distributed over time consolidate memory far better than massed practice — one of the most reproduced effects in cognitive psychology.
In GraphChinese: the decay model above schedules every review. You review a word or pattern when your modeled memory of it is about to dip, not when a deck happens to come up.
9. Interleaving
Mixing topics within a practice session beats blocked repetition of one skill, because you also learn to match each problem to the right technique.
In GraphChinese: reviews round-robin across domains — never three items from the same topic in a row — and a timed quiz every 150 XP mixes everything you've learned so far.
10. The testing effect
Retrieval is not just measurement; the act of recalling is the act of learning (Roediger & Karpicke's studies are the modern touchstone).
In GraphChinese: no reference material during reviews or quizzes — the lesson is hidden. "I'm totally stuck" re-queues the exercise without mastery credit rather than revealing the answer. And if an exercise could be answered from context by someone who never learned the topic, it gets rewritten.
11. Gamification, aligned
Points and streaks help only when they reward learning and can't be farmed without it.
In GraphChinese: XP and the daily goal are earned by mastery evidence — genuine retrieval — never by button-mashing. The daily goal is a pace, not a gate; nothing locks if life happens.
Why flashcard apps stall
Anki gets spaced repetition right and nothing else. A flashcard deck reviews words, but Chinese is a lattice of skills: tones support pinyin, pinyin supports characters, characters support words, words support grammar patterns, and grammar patterns support the sentences you actually say. A deck has no dependency graph, so it can't know that your shaky 了 usage traces back to a pattern you never mastered — it just keeps showing you cards. GraphChinese schedules reviews the same way Anki does, then adds the part Anki structurally can't: the graph that decides what you're ready to learn next and what a miss actually means.
Streak-first apps have the opposite problem: the mechanics reward showing up, not mastering anything, so years of daily use can leave someone unable to hold a two-minute conversation. Every gamified mechanic in GraphChinese is checked against one test — can it be farmed without learning? If yes, it's broken by definition.
What this looks like in a session
You open the study room and the selector hands you your frontier: maybe a review of 买 that's about to decay, a rapid-recognition drill on third-tone pairs, then a new grammar topic whose prerequisites you finished last week. The new lesson is a few short chunks with worked examples, each adding one element. You solve scaffolded exercises, then harder ones at genuine exam level — bite-sized never means watered down; an HSK 6 item here looks like real HSK 6. At 150 XP a ten-minute quiz interleaves everything, closed-book. Speaking is trained from lesson one with the comment-and-question method: memorize a small kit of comments and questions, and use them in real multi-round conversations long before your grammar is pretty.
Sources
- Justin Skycak & Math Academy, The Math Academy Way — the direct inspiration; GraphChinese adapts its mastery-learning engineering from math to Mandarin.
- Hermann Ebbinghaus, Über das Gedächtnis (1885) — the forgetting curve and the earliest spacing measurements.
- Benjamin Bloom, "The 2 Sigma Problem" (1984) — mastery learning and tutoring versus conventional instruction.
- John Sweller — cognitive load theory; instruction must fit working-memory limits.
- Nelson Cowan (2001) — the ~four-chunk capacity of working memory.
- Henry Roediger III & Jeffrey Karpicke (2006) — the testing effect: retrieval practice beats restudying.
- Cepeda, Pashler, Vul, Wixted & Rohrer (2006) — meta-analysis of distributed practice.
- Doug Rohrer — interleaved practice and its advantage over blocked practice.
Related guides: How to learn Chinese from absolute zero · The HSK levels, explained · The comment + question method
See it instead of reading about it. Your first lesson — the four tones, then tracing 你 and 好 in correct stroke order — is free on the landing page, no signup. The whole course, HSK 0 through HSK 6, is a one-time $49 for lifetime access.
Start the free first lesson