Remembering kanji is retrieval

Kanji stick when you produce them. They feel stuck when you reread them. Those are different events, and a 2008 vocabulary experiment measured the gap on a one-week delay. It did not use kanji. It used Swahili–English pairs. The schedule is what is worth copying. The language is not a claim that your deck will score 80%.

This is not a RubricMark trial. We did not run it. The paper is Karpicke and Roediger, Science, 2008.

What they actually varied

Students learned 40 Swahili–English pairs. Once a pair was recalled, the experiment either kept it in the study pile, kept testing it, or dropped it. Four endings:

• Study and test kept — about 80% a week later.
• Test kept, study dropped — about 80%.
• Study kept, tests dropped — 36%.
• Both dropped after one correct recall — 33%.

The two schedules that kept the test landed together. Extra reading, after you could already say the pair once, did not. The chart uses “about 80” as 80, which is how the paper summarises those two bars. The 36 and 33 are the reported figures for the schedules that stopped testing.

[[chart:ld-swahili-retrieval]]

A kanji card is the same shape as those pairs: a front you see, a back you must produce. The reading, the meaning, the stroke — whichever one you put on the back. Looking at a beautiful list of the grade-1 80 is the “study kept” arm. Covering the reading and writing the character is the test. The paper says keep the test in the pile after the first success. One correct recall is not graduation.

The same pattern, one week versus five minutes

An earlier prose study is the right caution about tonight’s mood. Roediger and Karpicke, 2006, Experiment 2: four quiet rereads recalled more at five minutes (83% of idea units) than a schedule with three tests (71%). One week later the rereads were at 40% and the repeated tests at 61%. Immediate ease is a bad forecast. Leave the chart on both delays, then switch to five minutes and to one week. The tallest bar changes side. That motion is the finding. It is not a kanji score.

[[chart:ld-testing-effect]]

A week of kanji that uses the result

Take the grade-1 80, or the next 15 of them if 80 is a fantasy for this week. Each card: character on the front, one reading and one word on the back. Session: cover, produce, rate. Leave the ones you got right in the pile. The experiment’s losing move was dropping them because they felt done.

A week that keeps the test

This is the schedule the Swahili result supports, applied to kanji without pretending the paper tested kanji.

1. Day 1, 25 min. Fifteen new cards from grade 1, or the next fifteen you cannot write. Study, then test. Anything missed once stays.
2. Day 2, 20 min. Test all fifteen before you teach yourself anything. Only then look at the back of the misses. The “study kept, tests dropped” arm is what you are refusing.
3. Day 4, 15 min. Test again. Add five new cards only if the old fifteen are mostly hits. New cards do not replace the test of the old ones.
4. Day 7, 15 min. Cold test, no warm-up reread. That score, not day-1 smoothness, is the one the papers are about.

Why this method works

The test arm and the study arm felt different in the room. Students who reread thought they knew the pairs. Students who kept being tested felt less fluent and remembered more a week later. Kanji lists are built to produce that illusion: the character is still on the page, so recognition pretends to be recall. Covering the page is the whole method. The grade list only decides which 15 characters go on the cards.

Run this session

Use the week above. If you only have one sitting, do day 1 and the day-2 test back to back: 25 minutes of new cards, a two-minute break, then 10 minutes in which you may not look at any back you got right. The break is there so the second pass is not echoic memory.

Rate the pile in Study Lab. When you want a timed paper, JLPT is the sit. Remembering the character is the part the sit will not do for you the night before.