Split View: 다시 읽지 말고 덮고 떠올려라 — 시험 효과의 원 논문과 간격 반복 알고리즘

Do Not Reread — Close the Book and Recall: The Original Testing-Effect Paper and the Spaced-Repetition Algorithm

Introduction — The Most Popular Study Method Is the One That Betrays You Most

Picture the night before an exam. Almost everyone does the same thing: rereading the textbook and their notes. They run a highlighter across the page, read the underlined parts again, and feel reassured by pages that already look familiar. "Now I know this."

That reassurance is the trap. What rereading builds is not memory but fluency — the ease of words sliding smoothly across your eyes — and the brain mistranslates that fluency as "I know this." This seventh installment of Psychology, Straight from the Papers digs into the paper that exposed this illusion experimentally, and into its alternative all the way down to the code. It will be the most practical installment in the series.

The 2006 Experiment — Rereading vs. Testing Yourself

The design of the 2006 paper by Henry Roediger and Jeffrey Karpicke (Psychological Science) is textbook-clean. College students studied short passages, but the way they studied differed.

• SSSS condition: read the passage four times (pure rereading)
• SSST condition: read three times, then close the book and write down everything recalled (one test)
• STTT condition: read once, then three consecutive recall tests (three tests)

And they split the final test across two time points: five minutes later, and one week later.

On the test five minutes later, rereading wins. But after a week the ranking flips completely. The group that read four times lost more than half, while the group that read once and retrieved three times kept most of it. Reading is a skill for the short game; retrieval is a skill for the long game.

There is a crueler detail. When students were asked right after studying to predict "how much will you remember a week from now," the rereading group predicted their own memory the highest — even though they were in fact the group that would forget the most. Feeling that you know and actually knowing are different systems, and rereading inflates only the former.

This is the testing effect. Before a test is an assessment tool, it is a learning tool. The very act of pulling something out of memory strengthens that memory's storage. The effect has been confirmed again and again in later classroom field studies and meta-analyses, and is one of the most trusted findings in the science of learning.

The Second Ingredient — The Spacing Effect

The partner of the testing effect is the spacing effect. It is the longest-surviving law of memory since Ebbinghaus's forgetting curve of 1885: for the same total amount of review, spacing it out (spaced) is overwhelmingly better for long-term memory than cramming it together (massed). The 2006 meta-analysis by Nicholas Cepeda's team (254 studies, more than 14,000 people) quantified this, and follow-up work even offered a practical rule of thumb: set your review interval to roughly 10-20% of the time left until the test (a test a month away means a 3-6 day interval).

Why is spacing effective? The leading explanation is "desirable difficulty." At the moment of review, the memory has to be slightly faded so that retrieval demands effort, and the more effortful the retrieval, the more strongly the memory is re-stored. Reviewing today a word you memorized yesterday is too easy; a month later it is already gone. There exists an optimal moment, right on the edge of being forgotten.

So can a person track the "optimal moment" for each of thousands of cards? No. That is where the algorithm comes in.

Understanding It Through Code — SM-2, the Heart of Anki

The prototype for spaced-repetition software (Anki and the like) is the SM-2 algorithm, which came out of Piotr Wozniak's SuperMemo in the 1980s. The principle is simple: keep a difficulty factor for each card, and every time you recall it successfully, stretch the next review interval by that factor. Fail, and the interval resets.

def sm2_review(quality, reps, interval, ease):
    """One SM-2 update after a review.
    quality: 0-5 self-grade (>=3 means recalled)
    reps: successful reviews in a row so far
    interval: current interval in days
    ease: difficulty factor, starts at 2.5
    Returns (reps, interval, ease) for scheduling the next review.
    """
    if quality < 3:                    # forgot -> relearn from scratch
        return 0, 1, ease

    # ease drifts with how hard the recall felt (min 1.3)
    ease = max(1.3, ease + 0.1 - (5 - quality) * (0.08 + (5 - quality) * 0.02))

    if reps == 0:
        interval = 1                   # first success: see it tomorrow
    elif reps == 1:
        interval = 6                   # second success: ~a week out
    else:
        interval = round(interval * ease)  # then multiply out
    return reps + 1, interval, ease

# simulate one card always recalled with quality 4
reps, interval, ease = 0, 0, 2.5
schedule = []
for _ in range(7):
    reps, interval, ease = sm2_review(4, reps, interval, ease)
    schedule.append(interval)
print(schedule)   # -> [1, 6, 15, 37, 90, 219, 533]

Look at the output. 1 day, 6 days, 15 days, 37 days, 90 days... the review interval widens exponentially. Seven reviews schedule a single card as far as a year and a half out. Twenty lines of code implement the spacing effect (an optimal interval that keeps widening) and the testing effect (every review is a retrieval test) at the same time. This exponential spacing is why adding 20 new cards a day does not make your daily review load explode.

For reference, modern Anki, on top of SM-2's descendants, also offers newer schedulers like FSRS that model recall probability explicitly. But whatever the scheduler, the heart is the same: retrieve it, right before you forget it.

What We Take Away — A Swap List for How You Study

The practice for this installment is clear: run the following swaps in your study routine.

• Rereading → close and recall. Once you have read a chapter, close the book and summarize it on a blank page. Wherever you get stuck is exactly what to review. It means choosing the discomfort of retrieval over the reassurance of fluency — the "outside the comfort zone" principle of deliberate practice applied to memory.
• Highlighting → writing questions. Underlining is scheduling a future reread. Instead, right there, write the questions you will ask yourself. A good list of questions becomes your exam paper as is.
• Cramming → a spaced schedule. If the test is four weeks away, do not pile the same 8 hours into one day; split it into four sessions at 3-5 day intervals. Same total, different result.
• Memory-heavy subjects → a spaced-repetition tool. Fact-dense areas like vocabulary, certifications, and medical or legal knowledge are the home turf of Anki-style tools. That said, cards you make yourself beat cards made by others — because making the card is itself the first retrieval practice.

It applies to developers just the same. The syntax of a new language, commands you keep forgetting, system-design patterns — because there is a final exam, the interview and the outage, where "I will just search for it" does not work, an engineer's studying is ultimately a game of retrieval too.

Reading Guide

• Testing effect: Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255.
• Spacing meta-analysis: Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354-380.
• Field-application review: Dunlosky, J., et al. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58. — A ranking table of 10 study techniques. The highlight is the scene where rereading and highlighting receive the lowest grade.

Reading tip: the Roediger and Karpicke paper puts everything in Figure 1 (a bar chart by condition for the 5-minute, 2-day, and 1-week tests). For the Dunlosky review, the overall grade-by-technique in Table 4 alone replaces ten study-method books. The next installment is the paper edition of the body series, the dose-response experiment on sleep debt.