# The Science of Efficient Kanji Acquisition: Five Evidence-Based Strategies

## The empirical case for structured kanji study

2,136 [joyo kanji](https://en.wikipedia.org/wiki/J%C5%8Dy%C5%8D_kanji) is not a willpower problem. It is a scheduling, encoding, and retrieval problem — and each of those three has decades of controlled experiments behind it. Five strategies below. Effect sizes, not anecdotes.

### 1. Spaced repetition — beat the forgetting curve with an algorithm

In 1885, [Hermann Ebbinghaus](https://en.wikipedia.org/wiki/Hermann_Ebbinghaus) sat down with lists of nonsense syllables — WID, ZOF, consonant-vowel-consonant trigrams designed to carry no meaning — and measured how much relearning he could skip at seven intervals. N=1. The subject was himself. The data has held up for 140 years.

| Time since learning | Savings (%) | Memory lost (%) |
|:---:|:---:|:---:|
| 20 minutes | 58.2 | 41.8 |
| 1 hour | 44.2 | 55.8 |
| 8.8 hours | 35.8 | 64.2 |
| 1 day | 33.7 | 66.3 |
| 2 days | 27.8 | 72.2 |
| 6 days | 25.4 | 74.6 |
| 31 days | 21.1 | 78.9 |

*Ebbinghaus, [Über das Gedächtnis](https://en.wikipedia.org/wiki/%C3%9Cber_das_Ged%C3%A4chtnis) (1885), §29. Replicated by [Murre & Dros (2015)](https://doi.org/10.1371/journal.pone.0120644), PLOS ONE.*

![Ebbinghaus forgetting curve with successive review cycles flattening the decay](https://upload.wikimedia.org/wikipedia/commons/4/4e/ForgettingCurve.svg)
*The [forgetting curve](https://en.wikipedia.org/wiki/Forgetting_curve). Red is unaided decay. Each green curve is a well-timed review flattening what comes next — the mechanical basis of spaced repetition. Source: [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:ForgettingCurve.svg).*

Ebbinghaus fit the decay to b = 100k / ((log t)^c + k), with k = 1.84, c = 1.25. Skip review and half the material is gone within an hour. A spaced repetition system exploits the curve by scheduling the next review at the moment recall probability slides toward a target threshold — late enough to make the retrieval effortful, early enough that you still succeed.

Algorithms differ in how aggressively they model the decay. The [open-spaced-repetition](https://github.com/open-spaced-repetition) benchmark — 727 million reviews from 10,000 Anki collections — lets you compare them on the same data:

| Algorithm | Model | Parameters | Log loss | Superiority vs SM-2 |
|-----------|-------|:---:|:---:|:---:|
| [Leitner](https://en.wikipedia.org/wiki/Leitner_system) | Box-based intervals | 0 | — | baseline |
| [SM-2](https://en.wikipedia.org/wiki/SuperMemo#Description_of_SM-2_algorithm) | Linear ease factor | 2 | 0.346 | — |
| HLR (Duolingo) | [Half-life regression](https://github.com/duolingo/halflearn) | 3 | 0.327 | — |
| [FSRS](https://github.com/open-spaced-repetition/fsrs4anki) v4 | DSR power-decay | 17 | 0.326 | 99.6% |
| FSRS-7 | DSR power-decay | 21 | 0.324 | 99.6% |

*[srs-benchmark](https://github.com/open-spaced-repetition/srs-benchmark), 10,000 collections. Log loss is cross-entropy between predicted recall probability and actual outcome. Lower is better.*

FSRS models memory with three state variables: **Stability** S (days for retrievability to drop to 90%), **Difficulty** D (1–10), and **Retrievability** R. The forgetting curve is a power function — R(t, S) = (1 + t/9S)^(-1). After each successful review, stability updates via S' = S · (1 + e^(w8) · (11 − D) · S^(-w9) · (e^(w10(1−R)) − 1) · hard_penalty · easy_bonus). Seventeen weights, fit per user via gradient descent over your own review history. My Kanji runs FSRS-5 natively in the [review system](/lessons/session).

### 2. Retrieval practice — the test is the encoding event

[Roediger and Karpicke (2006)](https://doi.org/10.1111/j.1467-9280.2006.01693.x) split 120 undergrads three ways on prose passages: four study periods (SSSS), three study plus one test (SSST), one study plus three tests (STTT). At five minutes, SSSS won (83% recall vs 71%). At one week, the order inverted: STTT held 61%, SSSS collapsed to 40%. The test was not measurement. The test was the encoding event.

[Rowland's 2014 meta-analysis](https://doi.org/10.1037/a0037559) of 159 comparisons pinned the [testing effect](https://en.wikipedia.org/wiki/Testing_effect) at Hedges' g = 0.50 (95% CI: 0.42–0.58). Medium-to-large, robust across conditions. For kanji, "active recall" means producing "water" when shown [水](/kanjis/6c34) — not reading "水 = water" forty times. Our [study sessions](/lessons/session) run four retrieval directions: meaning, reading, glyph-from-meaning, glyph-from-reading. The asymmetry is the point — recognizing a kanji and producing it are different skills, exercised separately.

### 3. Motor encoding — write it by hand

[Longcamp, Boucard, Gilhodes, and Velay (2006)](https://doi.org/10.1016/j.humov.2006.07.007) taught adults novel graphic characters by handwriting or typing, then tested recognition weeks later. Handwriting won — significantly stronger and longer-lasting recognition, and fMRI showed greater activation in the left [fusiform gyrus](https://en.wikipedia.org/wiki/Fusiform_gyrus) and inferior frontal regions. The same circuits the brain recruits during reading.

![Medial surface of the cerebral cortex with the fusiform gyrus highlighted](https://upload.wikimedia.org/wikipedia/commons/f/fd/Medial_surface_of_cerebral_cortex_-_fusiform_gyrus.png)
*Medial view of the cerebral cortex, fusiform gyrus highlighted. The left fusiform houses the [Visual Word Form Area](https://en.wikipedia.org/wiki/Visual_word_form_area). Longcamp et al. found it fires harder for characters learned by handwriting than by typing. Source: [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Medial_surface_of_cerebral_cortex_-_fusiform_gyrus.png).*

[Naka and Naoi (1995)](https://doi.org/10.3758/BF03197245) replicated the effect on Japanese materials. Repeated writing improved free recall of graphic forms over visual-only study, strongest when encoding and test modality matched. The mechanism: the motor trace of stroke production is a second, independent retrieval cue. A few correct-stroke-order repetitions per character buy you a memory channel pure-digital study cannot reach.

### 4. Interleaving — mix the confusable ones

[Kornell and Bjork (2008)](https://doi.org/10.1111/j.1467-9280.2008.02127.x) had participants learn painting styles from twelve artists. One group studied in blocks — six paintings by artist A, then six by B, etc. The other group saw them interleaved. On a transfer test with novel paintings, interleaving hit 59% accuracy; blocking, 36%. Effect size d = 0.99. Participants reported feeling blocking worked better. Their own data said otherwise — the metacognitive illusion is part of the finding.

For kanji this maps cleanly. Study visually adjacent characters in one mixed deck — [待](/kanjis/5f85) (wait), [持](/kanjis/6301) (hold), [特](/kanjis/7279) (special), [寺](/kanjis/5bfa) (temple) — and the [discrimination](https://en.wikipedia.org/wiki/Discrimination_learning) becomes the work. [Bjork and Bjork's](https://bjorklab.psych.ucla.edu/research/) [desirable difficulties](https://en.wikipedia.org/wiki/Desirable_difficulty) framework names the principle: conditions that slow initial acquisition (spacing, interleaving, retrieval) raise long-term retention by forcing deeper processing. The friction is the mechanism, not the obstacle.

### 5. Component decomposition — structure before memorization

[Atkinson and Raugh (1975)](https://doi.org/10.1037/0278-7393.1.2.126) tested the [keyword method](https://en.wikipedia.org/wiki/Mnemonic#Keyword_method) on Russian vocabulary: 72% recalled after one week with keyword mnemonics, 46% with rote. [Heisig's *Remembering the Kanji*](https://en.wikipedia.org/wiki/Remembering_the_Kanji_and_Remembering_the_Hanzi) (1977) generalizes the move to the full joyo set by decomposing each character into recurring "primitives" and linking them through narrative imagery. No head-to-head RCT on RTK vs traditional instruction at scale, but the keyword scaffolding underneath it is one of the better-attested findings in word acquisition.

The deeper point is compositional. Roughly 200 unique components — [radicals](/radicals), [graphemes](/graphemes), phonetic series — recombine across every joyo character. Learn [語](/kanjis/8a9e) as [言](/kanjis/8a00) + [五](/kanjis/4e94) + [口](/kanjis/53e3), not as a fourteen-stroke monolith, and the cost function turns from O(n) memorization into O(log n) structural analysis. Every component you add pays out across dozens of future characters. The [Kanji Atlas](/components) renders that graph for the entire joyo set.

### Key studies summary

| Study | Year | N | Effect size | Core finding |
|-------|:---:|:---:|:---:|-------------|
| [Ebbinghaus](https://en.wikipedia.org/wiki/Hermann_Ebbinghaus) | 1885 | 1 | — | Forgetting is power-law; 58% at 20 min, 21% at 31 days |
| [Atkinson & Raugh](https://doi.org/10.1037/0278-7393.1.2.126) | 1975 | 120 | 72% vs 46% | Keyword mnemonics beat rote at 1 week |
| [Naka & Naoi](https://doi.org/10.3758/BF03197245) | 1995 | — | sig. | Handwriting beats visual-only for graphic forms |
| [Roediger & Karpicke](https://doi.org/10.1111/j.1467-9280.2006.01693.x) | 2006 | 120 | 61% vs 40% | Three tests beat four study sessions at 1 week |
| [Longcamp et al.](https://doi.org/10.1016/j.humov.2006.07.007) | 2006 | — | sig. (fMRI) | Handwriting recruits reading circuits typing does not |
| [Kornell & Bjork](https://doi.org/10.1111/j.1467-9280.2008.02127.x) | 2008 | 120 | d = 0.99 | Interleaving doubles classification accuracy |
| [Rowland (meta)](https://doi.org/10.1037/a0037559) | 2014 | 159 studies | g = 0.50 | Testing effect is robust across conditions |
| [Ye (FSRS)](https://github.com/open-spaced-repetition/srs-benchmark) | 2023 | 10K users | 99.6% sup. | FSRS beats SM-2 on 99.6% of collections |

Five strategies, one stack. Decompose the character into components. Write it by hand. Test yourself in interleaved sets of look-alikes. Anchor it in a real word. Let an FSRS scheduler pick the next review. Each layer adds an independent memory trace, and any one of them is enough to pull the character back.

### References

- Atkinson, R.C. & Raugh, M.R. (1975). An application of the mnemonic keyword method to the acquisition of Russian vocabulary. *Journal of Experimental Psychology: Human Learning and Memory*, 1(2), 126–133. [doi:10.1037/0278-7393.1.2.126](https://doi.org/10.1037/0278-7393.1.2.126)
- Bjork, R.A. & Bjork, E.L. (2011). Creating desirable difficulties to enhance learning. In *Psychology and the Real World* (pp. 56–64). Worth Publishers. [PDF](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/04/EBjork_RBjork_2011.pdf)
- Ebbinghaus, H. (1885). *Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie*. Duncker & Humblot. English: *Memory: A Contribution to Experimental Psychology* (1913). [Archive.org](https://archive.org/details/memorycontributi00ebbiuoft)
- Kornell, N. & Bjork, R.A. (2008). Learning concepts and categories: Is spacing the "enemy of induction"? *Psychological Science*, 19(6), 585–592. [doi:10.1111/j.1467-9280.2008.02127.x](https://doi.org/10.1111/j.1467-9280.2008.02127.x)
- Longcamp, M., Boucard, C., Gilhodes, J.C., & Velay, J.L. (2006). Remembering the orientation of newly learned characters depends on the associated writing knowledge. *Human Movement Science*, 25(4–5), 646–656. [doi:10.1016/j.humov.2006.07.007](https://doi.org/10.1016/j.humov.2006.07.007)
- Murre, J.M.J. & Dros, J. (2015). Replication and analysis of Ebbinghaus' forgetting curve. *PLOS ONE*, 10(7), e0120644. [doi:10.1371/journal.pone.0120644](https://doi.org/10.1371/journal.pone.0120644)
- Naka, M. & Naoi, H. (1995). The effect of repeated writing on memory. *Memory & Cognition*, 23(4), 431–436. [doi:10.3758/BF03197245](https://doi.org/10.3758/BF03197245)
- Roediger, H.L. & Karpicke, J.D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. *Psychological Science*, 17(3), 249–255. [doi:10.1111/j.1467-9280.2006.01693.x](https://doi.org/10.1111/j.1467-9280.2006.01693.x)
- Rowland, C.A. (2014). The effect of testing versus restudy on retention: A meta-analytic review of the testing effect. *Psychological Bulletin*, 140(6), 1432–1463. [doi:10.1037/a0037559](https://doi.org/10.1037/a0037559)
- Ye, J. (2023). [FSRS4Anki](https://github.com/open-spaced-repetition/fsrs4anki). Benchmark: [open-spaced-repetition/srs-benchmark](https://github.com/open-spaced-repetition/srs-benchmark).