Skip to content

Retrieval Practice

Principle · All principles
Evidence · 134 claims (61 for, 63 mixed, 10 against) · 61 studies (29 causal, 14 review, 8 quant-synthesis, 4 associational, 3 qualitative, 3 theoretical), q1–q4 · 20 of 61 report an effect size · 127 claims rest on one study

Conditional relationship

For a learner who has already studied some material, recalling it from memory, rather than restudying it or not being tested, tends to raise performance on a delayed test of that material. The best-supported cases are adults and university students learning prose, word pairs and facts (verbal-association, concept), compared with restudy at a horizon of a day or more (delayed-retention). Two conditions bound the relationship. The initial retrieval has to mostly succeed, or be followed by corrective feedback; repeated failure with no feedback has shown about no benefit. And the horizon matters: on a test minutes after study, restudy has been ahead. Transfer to new questions (near-transfer) is smaller and depends on how the retrieval and the later task are configured. Whether the benefit holds for material high in element interactivity is contested. This is a bounded expectation for groups under stated conditions, not a forecast for one learner.

The Team-Based Learning pattern is one reusable design that puts this relationship to work: its individual readiness test is a retrieval attempt, and its team retest supplies immediate feedback. The retrieval evidence bears on those steps, not on the team formation, application exercises or peer evaluation around them. The pattern's response-dependent branches are design proposals, untested with learners.

Fitting the design to a situation

The rows draw on the study tables gathered under The evidence by question. Ask for three facts when the brief omits them: how long until the test that matters?, is the goal retention or transfer?, and will feedback be given?

If the brief says… Then change… Basis
The goal is retention of facts, terms or procedures over weeks or a course Replace some restudy or rereading time with recall: low-stakes quizzes or practice tests on the material just studied Retrieval practice improves long-term retention [+S]; Practice testing's meta-analytic benefit [+M]
The setting is a school or college classroom with unit and end-of-term exams Run short quizzes in class (clickers, software or paper), daily or at the end of each lecture, on the content the exams will cover. Where quizzes are voluntary, expect uptake to vary and record it: in two undergraduate courses, students who took more of the online practice tests scored higher on the exam with prior knowledge controlled, an association in self-selected use, not a comparison of assigned conditions Classroom quizzing improves classroom exam performance in middle school and college [+M], second-hand in one review chapter; More voluntary practice tests went with higher exam scores [+W]
The test that matters is the same day or the next morning Expect a smaller benefit, or none; restudy can match retrieval at very short delays. Judge the design on a test a day or more later Retention interval moderates the testing effect [~M]; Retrieval practice effects are larger at retention intervals over one day [~M] (Rowland's g = 0.69 against 0.41, reported second-hand)
The goal is transfer: inference, application, new problems Ask practice questions that require inference or application, not only the items to be remembered; expect a smaller effect than for retention. Do not expect practice tests in one subject to raise general problem-solving: in a randomised experiment with about 19,000 15-year-olds in 32 countries, more mathematics on a practice test raised mathematics scores by about 2% of a standard deviation per hour and did not change scores on non-curricular problem-solving tasks Retrieval practice improves transfer [+S]; Transfer is weakest to rearranged and untested material [~M]; More mathematics test practice gave a small mathematics gain [+M]; Mathematics test practice did not improve domain-general problem solving [-M]
Practice items can be answered wrongly with confidence, or use multiple choice Give corrective feedback after each attempt, and make the correction of a confident error explicit; without feedback, multiple-choice lures can be learned as facts, and more alternatives add to that risk when learners mostly fail the items. Feedback straight after the quiz or later in the session both serve: in one experiment either schedule brought lure intrusions down to the level of learners never shown the lures Feedback enhances retrieval practice [+S]; High-confidence errors are retained better after correction [+M]; Multiple-choice lures can be learned as false knowledge [-M]; More multiple-choice alternatives help at high initial success and hurt at low [~M]; Feedback on a multiple-choice test cut lure intrusions, immediate and delayed alike [+M]
Practice time per session is fixed (a timed app session, a flashcard block) and items are often answered correctly Let learners move past the feedback after a correct answer and spend the saved time on more recall attempts; keep the correction after an error. Undergraduates allowed to skip feedback made about six attempts per item instead of four and recalled more a day or two later. Untested beyond word pairs in a lab Skipping feedback in fixed time improved delayed recall [+M]; Skipping permits more retrieval attempts [+M]
The question is which item format to use Either format works; with feedback, short answer can do better, and without feedback multiple choice can do better because more items succeed. Matching the final test's format can help where it is known, but the benefit does not require an exact match. Choose the breadth of the task by what must be kept: short-answer questions on specific points retained those points better than free recall a week later, while free recall of a whole text helped the parts no question targeted, and only on a free-recall test (one within-subject study of 54 university students; overall scores did not differ) Targeted short-answer retrieval retained targeted content better [+M]; Free recall helped untargeted content, on a free-recall test only [~M]; Short-answer versus multiple-choice evidence is mixed [~M]; Feedback decides whether short answer beats multiple choice [~M]; Practice-test format moderates the effect [~M]; Benefits do not require matching initial and final formats [~M]
Learners are new to the material and would fail most recall attempts Give a brief study pass or cues first, so that most attempts succeed; then withdraw cues (free recall rather than recognition, fewer letter cues) as success rises, since less support during practice gave better retention when learners could still retrieve. Never give a multiple-choice quiz on material not yet studied without feedback: in one experiment learners then produced more lures than correct answers later Retrieval failure reduces benefit [~M]; Testing without study or feedback can be net harmful [~M]; Effects are more robust when initial retrieval success is high [+M]; Less initial retrieval support produces better retention [~M]
Practice can recur across the course Space the quizzes across sessions rather than massing them; repeated recall in a row right after study adds little, even though spaced attempts succeed less often during practice; an expanding schedule is not better than an even one Spaced retrieval improves retention [+M]; Spaced retrieval outperforms massed retrieval despite lower initial recall [+M]; Expanding and uniform schedules do not differ [~M]
Learners are school children, older adults, or the material is visual or spatial Use it; the benefit has been found across ages and materials, smaller in elementary school than at university, and clearer at about age 8 than age 6 in a few studies The testing effect generalizes across education levels [+M]; Benefits seen in children, older adults and memory-impaired patients [+M]; Benefits extend to nonverbal and spatial materials [+M]; Practice testing rated high utility across learners and contexts [+W]
Learners study on their own (online-self-paced, revision) Build the quizzes into the materials and tell learners why; left alone, most reread rather than test themselves. Where a system can schedule review, adapting it to each learner's responses did better than learner-directed study in one study; a fixed schedule did not differ reliably Learners underuse effective strategies [~M]; Learners misjudge the benefit of retrieval [~M]; Adaptive spaced retrieval beat learner-directed AI study at seven weeks [~M]
Learners are anxious about tests Keep quizzes low-stakes and ungraded; the benefit holds across anxiety levels but is smaller at high anxiety Higher anxiety is associated with smaller testing effects [~M]

Observation, state and explanation

Record a response under stated conditions: the item format (free recall, cued recall, short answer, multiple choice), cues and notes allowed, the time since study, the stakes, and whether feedback followed. A score on a quiz is an observation. "Has a durable, retrievable memory" is an inferred state, and so is "studied the reading". Correct recognition of an option is a different performance from producing the answer, and an answer given two minutes after reading says little about one given a week later. Confidence is a separate observation; learners who restudied in the cited experiment were more confident and did worse at a week, so confidence alone does not settle the state.

Strengthened memory access is the explanation the claim pages give for the effect. It is a mechanism hypothesis; no observation in a single quiz measures it. Two rival explanations have been tested and did not account for it. In word-pair experiments, attempting recall before each re-presentation beat study-only presentations of equal number (nonsense pairs [+M]); the advantage remained when only the last presentation differed, so a different way of encoding later presentations does not explain it (not an encoding strategy [+M]), and when each list was all one condition, so learners rehearsing the study items more does not explain it (not differential rehearsal [+M]). One laboratory study (Carrier & Pashler, 1992) with no effect sizes printed; it rules out alternatives, it does not show the mechanism.

Response Plausible explanations Observation that could change the interpretation
Recalls well right after studying, poorly a week later Fluent restudy created short-lived access; the material was never well encoded; later learning interfered Compare an unaided recall probe at the delay with the immediate one, on matched items; ask for a prediction of later recall first, and record whether any retrieval or feedback came between.
Fails most items on a first unaided attempt The material was not studied or not encoded; the cues on the probe do not match how it was learned; the stakes or anxiety suppressed responding Re-present the material, then probe again after a short gap; compare a cued with a free-recall version; compare a low-stakes attempt with a graded one.
Correct on multiple choice, wrong on short answer for the same content Recognition without recall; a lure picked earlier has been learned as the answer; unfamiliarity with the response format Give a short-answer version of the items; check whether wrong answers repeat options the learner selected before; note whether feedback followed the earlier test.
Recalls the fact, cannot use it on an application question The retrieved item is not connected to the principle the question needs; the question demands knowledge that was never practised; reading or language load Ask the application question with the fact supplied and without it; ask the learner to explain the relation; vary the wording of the question.
Scores rise across repeated quizzes but not on the later exam Item-specific memory of the quiz questions; the exam measures something else; the gap before the exam was longer than any practised one Probe with new items on the quizzed topics, at a delay like the exam's; compare the quiz and exam blueprints.

These are proposals for telling explanations apart, untested with learners. A probe is itself a retrieval attempt and can change what it measures; record it as an exposure. One contrast shifts confidence among the explanations; it does not identify a cause.

Evidence and qualifications

  • Retrieval practice improves long-term retention [+S]: two meta-analyses. The first, of 159 effect sizes from 61 studies, compared tested with restudied material (g = 0.50, 95% CI 0.42 to 0.58, heterogeneity high); the effect was larger at retention intervals of a day or more (0.69) than below a day (0.41), with feedback (0.73 against 0.39) and with cued recall rather than recognition (0.61 against 0.29), and about zero (0.03) with no feedback and initial success at or below 50%. Published studies showed larger effects than unpublished ones (0.58 against 0.25). The second, of practice testing across education levels, found g = 0.51, with multiple-choice and constructed-response practice tests both effective; its comparators mix rereading, filler tasks and no treatment, which is not the same contrast as testing against restudy. An experiment with undergraduates reading science passages found restudy ahead at five minutes (81% against 75%) and testing ahead at two days and one week; the claim page records it with two sample sizes (n=180, and 120 + 180 across two experiments), so read the full entry before quoting one. Not settled: the abstracts available could not confirm the entries. It does not predict an individual's gain or the size of an effect on complex skills.
  • Feedback Enhances Retrieval Practice [+S]: in a review of classroom studies (222 in the published abstract; 48,478 students), quizzing beat restudy and other controls overall (g = 0.499), and quizzes with corrective feedback produced g = 0.537 against 0.374 without. That is a between-study comparison, and another meta-analysis found no moderation (0.63 against 0.60). A word-pair experiment (258 participants, Luganda–English, final test one week later) found that supplying the answer after an error raised retention, while feedback after a correct response made little difference. One of two entries passes the judge; the other could not be confirmed. On timing, one multiple-choice experiment (72 participants) found immediate and delayed feedback equally effective at removing lure intrusions, bringing them down to the level of a group never shown the lures (d = .65 and .81 against no feedback) (lure intrusions [+M]); that is one outcome of one study, and the wiki records no comparison of timing on correct recall. Feedback after a correct answer has a cost when time is fixed: in a Swahili–English word-pair experiment, letting learners skip it bought more recall attempts (5.88 against 4 per item) and higher recall after 24–48 hours (.72 against .63) (skipping feedback [+M]; more attempts [+M]). Elaborated against answer-only feedback is not tested here. Errors made with high confidence are corrected more durably than low-confidence ones (hypercorrection [+M]; a narrative review and one experiment, no effect size printed, not confirmed against the sources). Without feedback, multiple choice carries a specific risk. The same 72-participant experiment found a multiple-choice test without feedback still beat no test on later cued recall (d = .97), but for items the learners had not studied first, lure intrusions (.32) outnumbered correct answers (.18), so testing unstudied material without feedback did harm (net harm boundary [~M]). And students later produced lures they had selected (lures [-M]), and more alternatives helped retention when initial success was high but hurt it when success was low (alternatives [~M]); both are reported second-hand in one review chapter.
  • Retrieval Failure Reduces Benefit [~M]: in the same 61-study meta-analysis, no-feedback studies whose initial recall was at or below 50% showed g = 0.03, and those above 75% showed g = 0.56. This is a moderator analysis across groups of studies, not a manipulation of success, so it gives no threshold for one learner. Not settled on the abstract. Pretesting studies, where failure followed by feedback helps, are a counterpoint the page records. The same review chapter reports that effects grow more robust as initial success rises, especially above 75% (success [+M]), and also that less support during practice helps when learners can still retrieve: free recall beat recognition although 89% of words were recognised and 69% recalled, and fewer letter cues beat more (less support [~M]). Read together, success and effort pull in opposite directions; the chapter gives no point at which they balance, and none of these were manipulations of success within one study.
  • Spacing the retrieval: a meta-analysis of 29 studies found spaced retrieval ahead of massed retrieval (g = 0.74) and no reliable difference between expanding and uniform schedules (spaced retrieval [+M]; not yet checked against its source). Experiments reported in the review chapter found that a small spacing cut initial recall by about 25% yet raised final retention, and that massed repeated retrieval gave no benefit at one week over dropping an item after one recall (spaced against massed [+M]). In one study, adaptive spaced retrieval beat learner-directed generative-AI study on a seven-week posttest (partial η² = .128), while a fixed spaced schedule did not differ reliably from either (adaptive retrieval [~M]); the comparator is learner-directed study, not restudy, and the study is a single preprint.
  • Settings and learners: classroom quizzing by clickers, software or paper improved classroom exams in eighth-grade science, college statistics and introductory psychology, including on new questions in one study (classroom quizzing [+M]); benefits are reported for children aged 9–11, adults aged 55–65 and memory-impaired patient groups, more evident at about age 8 than age 6 (learners [+M]); and higher trait or induced anxiety went with smaller, still positive effects (anxiety [~M]). All three are second-hand from one review chapter, with no effect sizes printed. In two undergraduate courses with four voluntary online practice tests, each test taken went with a higher exam score after prior knowledge was controlled (β = .27 without feedback, N = 201; β = .28 with corrective feedback, n = 111) (voluntary practice tests [+W]); students chose how many to take, so motivation or study time may carry part of that association. A monograph's narrative synthesis rates practice testing high utility across ages, abilities, tasks and contexts (utility [+W]), a summary judgment rather than a new comparison.
  • Format matching: the review chapter concludes that the benefit does not depend on an exact match between practice and final-test format (cue direction, map matching to route recall) (format matching [~M]), and that short answer beats multiple choice mainly when feedback follows, with multiple choice ahead without feedback in Kang et al.'s first experiment (86% against 54% initially correct) (feedback and format [~M]); across four experiments the format difference was d = 0.07 (format evidence mixed [~M]). Breadth is a separate choice from format: in a within-subject study of 54 university students reading two expository texts, short-answer practice on specific sections retained those sections better a week later on both posttest types (η² = .195 on the short-answer test), while free recall of the whole text retained the untargeted sections better only on a free-recall posttest (η² = .084) (targeted [+M]; holistic [~M]); overall scores did not differ (no overall difference [~W]).
  • Retrieval Practice Improves Transfer [~S]: a meta-analysis of 122 experiments (192 effect sizes, N = 10,382) found transfer from practice testing against non-testing re-exposure (d = 0.40, 95% CI 0.31 to 0.50), strongest to application and inference questions and across test formats, weakest to rearranged items, untested material and worked-example problems; bias corrections often left no positive transfer when the favourable moderators were absent. Four prose-passage experiments found repeated testing ahead of repeated study on new inferential questions a week later. Both entries pass the judge (abstracts). It does not establish far transfer to new problem structures, and one large field experiment points against far transfer from test practice: within PISA 2012, students randomly given forms with more mathematics content gained about 2% of a standard deviation per hour on a later mathematics task (b = 2.29, SE = 0.84; N = 19,355) and nothing measurable on non-curricular problem-solving tasks (b = 0.36, SE = 1.81; N = 19,481) (mathematics gain [+M]; no general problem-solving gain [-M]). The comparator was less mathematics within the same two-hour test, and the claim pages record nothing about feedback, so it measures more practice testing in a subject against less, not testing against restudy.
  • The argument that retrieval practice effects do not occur with materials high in element interactivity is contested [~M]: second-hand, from one review chapter: one study found effects with a randomly ordered text but not an intact one, another found them for both, and the chapter judges the limiting argument not convincing. Not yet checked against its sources. For high element-interactivity material or a complex-skill goal, treat the relationship as unsettled.

Retain the comparator (restudy, rereading, filler or nothing), the item format, feedback, initial success, the horizon and the learners when transporting these results. The pooled values come from different comparisons and should not be ranked against one another or read as an individual's expected gain.

The evidence by question

Each question below links a general claim that gathers the studies behind it into one table, one row per distinct study, with each study's design and effect as recorded and no averaged verdict (eval/synthesis/retrieval-practice.json, scripts/synthesize_claim.py; a pilot, 2026-10-13). The count is of distinct studies across the claim and the narrower claims that bear on it.

Further evidence, not yet read against this model

None at present. The sixteen restored claims were read on 2026-10-15; the eleven attached by the linker between 2026-10-15 and 2026-10-19 were read on 2026-10-19 and all bear on the model, so each now sits where it informs it: Butler & Roediger (2008) in the multiple-choice and novice rows, the feedback qualification and the feedback decision; Carrier & Pashler (1992) under Observation, state and explanation, as tests of rival explanations; Hays et al. (2010) in a new fixed-time row and the feedback decision; Endres et al. (2020) in the format row and decision; Naujoks et al. (2022) in the classroom row and the settings qualification; Avvisati & Borgonovi (2020) in the transfer row, qualification and decision. - No testing effect was found for acquiring problem-solving skills through worked example study across eight experiments [~W] — attached 2026-10-21 from van Gog et al. (2015), which proposed "Consider the complexity of learning materials when deciding between practice testing and further study"; tests this page's relationship. - Testing effect disappears for a coherent complex text but reappears, as reduced forgetting, when the same sentences are scrambled [~W] — attached 2026-10-21 from van Gog et al. (2015), which proposed "Consider the complexity of learning materials when deciding between practice testing and further study"; tests this page's relationship. - The advantage of recitation over rereading is greater the less connected the learning material (Kühn 1914; Gates 1917) [+W] — attached 2026-10-21 from van Gog et al. (2015), which proposed "Consider the complexity of learning materials when deciding between practice testing and further study"; tests this page's relationship. - Overt retrieval practice produces higher delayed recall of definitions than covert retrieval practice for learners studying declarative concepts [+M] — attached 2026-10-21 from Michelle L. Rivers et al. (2025), which proposed "Students should prioritize overt retrieval when mastering complex information such as definitions". - Covert retrieval is the most time-efficient practice strategy, yielding the largest gains per minute of practice for both terms and definitions [~M] — attached 2026-10-21 from Michelle L. Rivers et al. (2025), which proposed "Students should prioritize overt retrieval when mastering complex information such as definitions". - For incoherent text presented as a list of facts, testing slows the rate of forgetting compared to restudy [+W] — attached 2026-10-22 from de Jonge et al. (2015), which proposed "Use completion tests for retention of unconnected materials rather than highly structured texts"; tests this page's relationship. - For a highly coherent text, taking a fill-in-the-blank practice test yields no retention benefit and does not slow forgetting compared to restudy [~W] — attached 2026-10-22 from de Jonge et al. (2015), which proposed "Use completion tests for retention of unconnected materials rather than highly structured texts"; tests this page's relationship. - Without initial instruction, varied retrieval practice matches varied worked examples and beats repeated retrieval practice for generalization [+M] — attached 2026-10-22 from Meng Cao et al. (2026), which proposed "Align learning strategy choice with the instructional goal: induction versus memory consolidation"; tests this page's relationship. - Images that reveal the answer during retrieval practice reduce vocabulary recall compared with practice without images [~W] — attached 2026-10-24 from van den Broek et al. (2021), which proposed "Avoid informative images during vocabulary retrieval practice because they can harm learning". - Stepwise MCQs, which delay answer options by 4 s, enhance vocabulary retention significantly compared to standard MCQs in three of four experiments, measured days after practice. [+W] — attached 2026-10-24 from van den Broek et al. (2023), which proposed "Use a stepwise display in MCQ-based retrieval practice: present the question before the answer options to create a cued-recall opportunity.". - The stepwise-MCQ benefit is most robust for questions with noncompetitive answer options and for learners who report trying to retrieve the answer during the delay. [+W] — attached 2026-10-24 from van den Broek et al. (2023), which proposed "Use a stepwise display in MCQ-based retrieval practice: present the question before the answer options to create a cued-recall opportunity.".

Objective and learner-valued goal

Ask what the learner wants to be able to do with the material and why, and record it apart from the designer's objective. Agreement, divergence and uncertainty are observations, not assumptions. Then fix the target: the capability (recall a term, explain a relation, apply it to a case), a representative instrument, a criterion chosen locally, the aids permitted, and the horizon. Retrieval evidence is mostly about delayed recall of studied material; if the target is application or a whole skill, say so, because the expectation is weaker there.

For example, a nursing student may value being able to answer a patient's question about a medication on placement, while the designer's objective is one-week cued recall of drug classes and their actions. A recall quiz serves the designer's objective directly and the learner's only in part: answering a patient is an application to a new question, which is where the transfer evidence is conditional. Keep both aims, and assess the application at its own horizon rather than inferring it from quiz scores.

What would revise this model?

The delayed-retention expectation should weaken if comparable learners, at equal time, on aligned delayed outcomes, show no advantage for retrieval over restudy in defensible comparisons, including unpublished ones; the publication-bias gap already recorded is a reason to watch for this. If within-study manipulations of feedback or of initial success fail to reproduce the between-study moderators, the conditions above should be restated. If retrieval on high element-interactivity material or complex-skill goals is shown, under controlled comparisons, to help or not to help, replace the "contested" line with that finding. Do not protect the model by relabelling every null as too little feedback or too low success after the fact.

A within-session quiz score, delayed recall, transfer to new questions and valued use are separate claims. The present evidence does not establish an optimal success rate for an individual, an optimal schedule, or a benefit across all domains.

Design Decisions

Should learners retrieve, or restudy?

Should feedback follow retrieval?

How hard should retrieval be?

Which question format: multiple choice or short answer?

When should retrieval happen?

Will it help learners apply knowledge, not only recall it?

For whom, and with what material?

Claims

Examples

Key Sources

  • Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
  • Karpicke, J. D. (2017). Retrieval-Based Learning: A Decade of Progress. Learning and Memory: A Comprehensive Reference, 2nd edition, Volume 2. https://doi.org/10.1016/B978-0-12-809324-5.21055-9
Page metadata
Field Value
type principle
id retrieval-practice
aliases balance-retrieval-success-and-retrieval-effort
title Retrieval Practice
description For learners who have studied material, recalling it from memory rather than restudying tends to raise delayed retention when initial retrieval mostly succeeds or is corrected by feedback; restudy can lead at a few minutes, transfer gains are smaller and conditional, and high element-interactivity material is contested.
status review
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 2 — roediger-2006, karpicke-2017

Frontmatter is the page's machine-readable record — see the Schema & Guide.