Your kid forgot 書 by Tuesday. That's fixable.

My daughter learned 書 last Tuesday at Saturday school. By Tuesday — the next Tuesday — it was gone. We’d practiced over the weekend. Her teacher had drilled it. Monday she could still pull it. Tuesday, nothing.

This is not a focus problem. It is not a sign she doesn’t care. It is how memory works, for everyone, in every subject, at every age. The fix has been understood for over a hundred years, and it is not expensive.

In 1885 a German psychologist named Hermann Ebbinghaus memorized nonsense syllables and tracked how fast he forgot them. Retention dropped to roughly half within an hour. Meaningful material — characters tied to words a heritage kid already says at dinner — fades slower, but the shape of the curve is the same. Without review, what you learned on Saturday is mostly gone by the following weekend. Cognitive scientists call this the forgetting curve, and it is one of the most replicated findings in the field.

I am the founder of MandarinBytes. Before that I spent a decade at Disney building personalization systems, and before that I was the kid in the back of Chinese school who forgot everything by Wednesday. My eight-year-old goes to Taiwanese Saturday school in Seattle. I have watched the forgetting curve operate on her in real time, with full knowledge of what it was and exactly zero ability to stop it by sheer parental will.

So I built the tool I wanted.

How reviewing at the right time changes everything

The trick is timing. Review a character five minutes after she learned it and the review does almost nothing — the memory is still warm, she’s recognizing the shape, not retrieving it. Wait a month and the memory is gone; she’s not reviewing, she’s starting over.

The useful window sits between those two. Reviewing right around the point she’d otherwise start to forget forces the brain to do real retrieval work, and that slight effort — pulling the character back from the edge — is what makes the next memory more durable. Each successful retrieval pushes the next forgetting point further out. One day. Three days. A week. A month. The intervals grow because the memory is actually getting stronger, not because the algorithm is being generous.

This is not a theory that sounds clean and falls apart in the field. Settles and Meeder, working with Duolingo’s data in 2016, found vocabulary retention improved measurably under spaced-repetition scheduling compared to fixed-interval review. Pimsleur proposed graduated-interval recall in 1967 and the basic shape has held up across sixty years of replication.

What the algorithm actually does

She reviews 書. The system records whether she got it, and how fast. From that it picks the next review date — late enough to be effortful, soon enough that the character isn’t gone.

Characters she nails get pushed out. After five clean retrievals 書 might not appear again for three weeks. Characters she fumbles come back tomorrow, or the day after. The scheduler is not running “a review session.” It is managing several hundred memory timelines in parallel, each on its own clock, each responding to her actual performance on that character.

The practical effect is that practice time lands where it matters. She is not burning ten minutes on characters that are rock-solid, and she is not losing 書 to the gap between Saturdays because that week happened to be busy.

No parent sorts flashcards. No teacher decides what tonight’s deck should be. The scheduler handles it.

Why this works especially well for heritage learners

Here is the thing most learning apps miss, and it is the entire reason I am writing this.

My daughter already knows the word “shū.” She has said it about a thousand times — at bedtime, at the library, when she’s hauling a stack of them off the shelf. The character 書 is not a new concept for her. It is a new shape attached to a sound and a meaning she has carried since she was two.

A kid learning Mandarin from scratch sees 書 and has to build three things at once: a shape, a sound, and a meaning, with no prior anchor. A heritage kid sees 書 and only has to build one mapping — the shape connects to the audio that was already there. That is a structurally easier problem. Not a little easier. Materially easier.

Spaced repetition exploits this. The memory isn’t being built on bare ground; it’s being attached to a network of meaning and sound she already carries from the dinner table. The heritage-language research is consistent on this point — prior spoken knowledge, when properly activated, accelerates literacy development. The intervals can stretch sooner because what’s underneath is more stable.

The kids with the most to gain from a well-scheduled review of 書 are the kids who already know what 書 means. They are also, somehow, the group nobody has built the tool for.

What FSRS is, briefly

The scheduler under the hood is called FSRS — Free Spaced Repetition Scheduler — developed by Jarrett Ye and collaborators, currently at version 4.5. Anki, the most widely used flashcard app in the world, ships it as the modern default. It replaced an older fixed-table approach with something that learns each user’s actual forgetting rate and schedules around that.

What FSRS does that earlier algorithms didn’t: it models your kid specifically. Two children reviewing 書 get different schedules based on their own history of right and wrong answers. The model gets sharper the more data it sees on a particular learner.

The thing I cared about, building MandarinBytes, is that it runs per skill per character. 大 isn’t one number in our system; it is three. My daughter knows 大 fluently by ear. She is still shaky on it in writing. One stability score across both would smooth that gap into a lie. So 大-by-listening might be scheduled thirty days out while 大-by-reading comes back tomorrow, and that is the honest shape of where she is.

What spaced repetition doesn’t do

Spaced repetition is not a fluency machine. It builds recognition and recall — your kid will see 書 and know it means book and sounds like shū, and that knowledge will still be there in March. That is the entire job, and it is a big job, because without it the Saturday-school cycle just resets every week.

What it does not do: teach your kid to hold a conversation, narrate a story, or use 書 in a sentence she made up. Those are different skills. They are built by talking at dinner, reading together on the couch, and the messy real-world language use that no algorithm has ever replaced and probably never will.

Treat SRS the way you’d treat a multivitamin. It is not the meal. It makes sure the foundation under the meal does not quietly crumble between Saturdays. The meal is the Mandarin you already speak at home, the books on your shelf, the cousin on FaceTime from Taipei. SRS keeps the characters your kid learns from leaking out of her head between exposures to those things.

If a tool promises more than that, it is selling you something.

The scheduling problem

Here is what I believe, having read the literature and watched my own kid forget 書: the heritage Mandarin attrition problem is not a motivation problem and it is not a parenting problem. It is a scheduling problem. The characters arrive faster than untutored memory can hold them, and Saturday school is structurally incapable of solving this on its own — there are not enough Saturdays.

An algorithm that schedules a two-minute review of 書 on the exact Wednesday your kid was about to lose it is not a gimmick. It is the thing that has been missing. Ebbinghaus published the forgetting curve in 1885. Pimsleur published graduated-interval recall in 1967. Anki has shipped a working spaced repetition scheduler since 2006. Heritage Mandarin learners — the kids with the most to gain from retention support, because they already have the spoken language and only need the writing system to stick — have somehow been the last group to get a tool built around any of this.

The forgetting curve is real and it does not care about your weekend plans. The fix is a hundred and forty years old. The question is only whether your kid’s review of 書 happens on the right Wednesday.

Research cited in this article

Ebbinghaus, H. (1885/1913). Memory: A Contribution to Experimental Psychology. — Original forgetting curve research; theoretical basis for spaced repetition.
Pimsleur, P. (1967). A memory schedule. Modern Language Journal, 51(2), 73—79. — Graduated-interval recall; early spaced repetition framework for language learning.
Settles, B. & Meeder, B. (2016). A trainable spaced repetition model for language learning. ACL 2016. — Empirical evidence for SRS retention advantages in language learning.
Ye, J., Tang, L., et al. (2024). FSRS-4.5: A Spaced Repetition Algorithm. — The scheduling algorithm discussed in this article.
Laufer, B. & Goldstein, Z. (2004). Testing vocabulary knowledge: Size, strength, and computer adaptiveness. Language Learning, 54(3), 399—436. — Receptive and productive vocabulary knowledge are dissociable.
Webb, S. (2008). Receptive and productive vocabulary sizes of L2 learners. Studies in Second Language Acquisition, 30(1), 79—95. — Productive vocabulary consistently smaller than receptive across proficiency levels.

MandarinBytes runs FSRS per skill per character, because listening and reading erode at different rates and one number across both is a lie. Reviews land on the days a real family is awake, not on a streak counter.