Retrieval, Spacing and Interleaving

Retrieval, spacing and interleaving — the three concrete study techniques that make practice effortful and, as a result, make what is learned durable. Retrieval practice asks the learner to produce an answer from memory rather than recognise one; spacing distributes that practice across time instead of massing it into one session; interleaving mixes problem types instead of blocking them. What unites them is a shared signature: each lowers performance during practice while raising retention after it — the performance–learning gap in operational form. They are the workhorses of Cognitive Psychology, and in AI-supported learning they are the specific behaviours that generative AI most easily erases, because an LLM that answers on demand removes the retrieval, flattens the schedule, and smooths away the mixing. This page is the techniques page. The principle behind them — Bjork's effort–learning trade-off, why harder practice produces more durable learning — is treated on Desirable Difficulties; this page covers what each technique is, what the corpus's studies measured, and how AI implements or undermines them.

Questions to Consider

• Have you ever re-read a chapter until it felt familiar, then been unable to explain it a week later? What did the re-reading actually give you?
• Retrieval practice means answering from memory before checking. When an AI assistant is one keystroke away, what would have to be true about your study routine for that retrieval to still happen?
• Spaced repetition systems schedule review at intervals a learner would not choose. Is a schedule that feels wrong a feature or a bug?
• Interleaving mixes problem types and feels messier than blocking. If the corpus offers almost no direct AI evidence on interleaving, how confident should you be in adopting it — and on what basis would you decide?
• Akgun and Toker (2026) found that students who chatted freely with an AI scored worst, even after the best pretesting. What did the free-chatting condition lack that the structured one had?
• Tsiligkiris (2026) found deeper LLM questioning improved task quality but not recall. What does that tell you about using fluency as evidence of learning?

Introduction

Retrieval practice, spacing and interleaving are usually taught and cited together because they were discovered separately but behave alike: each sacrifices how good practice looks in exchange for how much survives. In the knowledge base they appear as instances of the Desirable Difficulties principle, but they are also independently actionable — a reviewer can schedule intervals, a tutor can withhold the answer, a curriculum can mix item types, without invoking the underlying theory. Their operational character is why they are the natural bridge between Cognitive Psychology and AI design, and why the studies below tend to report two outcome measures rather than one: performance during practice, and retention after a delay.

The Testing Effect: Retrieval Practice Beats Re-Reading

Retrieval practice — attempting to recall material rather than re-reading it — strengthens memory more than additional study does, and the corpus treats the finding as settled enough to build on rather than re-litigate. The learning-to-learn scoping review places retrieval practice in the Tools layer of its three-layered framework (Dimensions of cognitive and metacognitive skill, Processes of self-regulation, Tools such as retrieval practice), positioning it as a concrete, teachable component of learning-to-learn that is meant to counterbalance GenAI overreliance and protect learner agency.

The most informative test in the corpus is a dissociation. Tsiligkiris (2026) logged 22 postgraduate students' prompt-by-prompt Large Language Models (LLMs) interactions during a neuroeconomics case task and separated depth (the proportion of explanation-seeking "why/how/explain" prompts) from volume and pacing. Depth predicted independently marked task quality (β = 6.27, p = .006) but had a null association with immediate post-test recall (β = −0.014, p = .728), with recall gains driven by baseline knowledge instead. The reading the author gives is that elaboration drives comprehension while retrieval drives consolidation — and that in LLM-supported study without explicit retrieval demands, learners can experience high fluency with little need to retrieve anything unaided. That is a retrieval-practice finding stated as an absence: the mechanism was not engaged, so the retention outcome did not move.

Retrieval practice is not a universal lever, and the corpus says so. Rachatasumrit, Koedinger and Carvalho (2025) ran a 2×2 experiment with 95 participants crossing knowledge content (verbatim facts vs. generalizable skills, geometry-area materials) with training schedule. They found a content–treatment interaction (β = 0.41, p = .038, d = 0.38): pure practice testing produced higher learning gains for facts, while example-integrated practice produced higher gains for skills. The authors also note that retrieval-practice gains frequently fail to extend to unfamiliar problems. So retrieval practice is strongest where the target is memory for specific content, and it needs to be paired with worked examples where the target is a generalizable skill — which is also why their AI design implication is that Intelligent Tutoring should adapt the example-to-practice ratio to the knowledge component being learned.

The same logic runs through the cheating-sheet study: Chen, Sakhnini and Istead treat constructing a cheat sheet as an active, generative study strategy (selection, condensation, organisation) rather than logistics, and frame the AI-era risk as offloading the artefact's construction and losing the metacognitive rehearsal that building it provided.

Spacing and Distributed Practice

The spacing effect — that the same total practice produces more durable learning when distributed across sessions than massed into one — is the most algorithmically exploited of the three techniques, and the corpus's applied systems are built around it.

Memdora is the clearest example of spacing treated as an engineering problem. It grounds scheduling in the Ebbinghaus forgetting curve, cites the figure that roughly 70% of newly learned material is forgotten within 24 hours without review, and integrates FSRS-6, described as the current state-of-the-art spaced-repetition algorithm. Its argument is that scheduling alone is not enough: existing tools reduce flashcard interaction to a single binary gesture, "flip and self-rate", which the authors call an impoverished model that fails to exploit cognitive-science evidence on retrieval practice. Memdora's contribution is therefore a taxonomy of 17 cognitively-grounded interaction types across Language, By Heart and Exam categories, each mapped to peer-reviewed evidence displayed on the card, plus an effort-based reward system that compensates actual cognitive engagement rather than app presence, a unified generation pipeline that creates cards at the point of reading, and a classroom layer that reports learning outcomes at the individual card level.

Akgun and Toker (2026) supply the corpus's strongest experimental evidence on how spacing is structured. In a three-arm randomized study of 89 undergraduates in an applied statistics course, all three groups practised in spaced sessions with identical number and timing over seven weeks; only the interaction structure differed. Adaptive spaced retrieval (G1) reached the highest posttest score (M = 78.19) and the highest observed practice effort (M = 0.85); fixed spaced retrieval (G2) followed (M = 74.55, effort 0.74); learner-directed AI study with no enforced retrieval came last on both (M = 67.28, effort 0.49). The multivariate effect was significant (Wilks' Λ = .664, partial η² = .185), with G1 ahead of G3 on retention at d = 0.92 (p = .003). The detail that matters pedagogically is that G1's agent was configured to withhold: response-contingent probes for Misconceptions about AI, requests for elaboration after superficial answers, advancement only on adequate conceptual engagement, and direct solutions explicitly excluded from its allowed outputs. The authors' own reading is that pretesting is a front-loaded catalyst rather than a standalone intervention whose benefits survive open-ended AI access.

Two further corpus findings sharpen what spacing does and does not yet get right. Schuetze, Yan and Carvalho (2025) fit Knowledge Tracing models to a six-session successive-relearning dataset and found that BKT, BKT-with-Forgetting and the Additive Factors Model reproduce learning trends when fit retrospectively (AUC 0.74–0.79) but, under time-based cross-validation of the kind a deployed tutor would need, overestimate future performance by roughly 58%, 51% and 47% respectively and fail to capture the spacing effect — sometimes predicting the opposite ordinal ordering across spacing conditions. In other words, the scheduling decisions made by adaptive systems may be based on models that do not represent the very effect spacing depends on. And LLM student modeling and long-term memory names the design gap directly: most tutoring systems lack longitudinal memory, and how such memory should interact with spaced repetition and forgetting curves remains an open question.

What the corpus does not yet show

Neither applied system has produced retention evidence of its own: the Memdora article reports improved retention relative to traditional spaced-repetition tools, but describes no delayed-interval outcome study behind that claim, and the bilingual lecture companion — which generates flashcards automatically to remove the authoring cost that blocks evidence-based retrieval and spaced practice, while simplifying scheduling to a binary "know it / still learning" rating rather than the graded SM-2 and FSRS algorithms — states plainly that no learning-outcomes study has been run, only a pre-registered protocol. The scheduling machinery is better evidenced than the learning it is meant to produce.

Interleaving

Interleaving — mixing problem or item types within a practice session rather than blocking them by type — is the weakest-evidenced of the three in this corpus, and that should be stated rather than padded. Interleaving appears on Desirable Difficulties only as one instance in a list of effortful conditions, and no study in the knowledge base tests an interleaving manipulation against a blocked control.

What the corpus does contain is adjacent and worth distinguishing carefully. Simulating learner task-selection models Interleaving and Blocking as two of eight candidate learner strategies in a Mastery Learning Simulation, alongside strength targeting, weakness targeting and outcome-informed rules. The finding there is about choice architecture rather than memory: some learner strategies systematically delay progression through over-practice, and task-selection constraints repaired the maladaptive strategies while leaving Interleaving, Blocking, Strength Targeting and Weakness Targeting at steady over-practice levels. That is evidence that interleaving is a selectable behaviour a system can accommodate or constrain — not evidence that interleaving improves retention. The remaining mentions in the corpus are unrelated senses of the word (interleaved agent stages in From Confusion to Consolidation: A Staged Conversational Workflow for Post-Lecture Review; interleaved visual-textual solution trajectories in a geometry Benchmark), and they should not be counted as interleaving evidence.

Interleaving's inclusion here therefore rests on the same theoretical family as the other two techniques and on its status in the Cognitive Psychology literature, not on a study in this knowledge base. Treat it as the open slot.

How AI Implements and Undermines These Techniques

Implementation. AI's most defensible contribution is scheduling and generation, the parts of these techniques that are tedious rather than pedagogically interesting. Adaptive scheduling is implemented in Memdora via FSRS-6 and in knowledge-tracing architectures that estimate mastery over time; adaptive pretesting is implemented by Akgun and Toker's response-contingent agent, which reads prior-session performance signals to raise conceptual depth in weaker areas while maintaining challenge in stronger ones; automatic flashcard and pretest generation is implemented in the bilingual lecture companion, where card authoring is the cost that blocks spaced practice in the first place. Longitudinal student memory is the architecture that would let any of this persist across semesters, and Woollaston and colleagues (2026) supply the design vocabulary for doing it deliberately: intentional friction, dynamic Scaffolding, human-in-the-loop oversight, and considered rather than maximal AI utilisation.

Undermining. The failure mode is specific and documented repeatedly. Woollaston et al. (2026) name the first of their six pedagogical risks as prior-knowledge activation: agents that pre-fetch and surface content bypass the retrieval practice that activates prior knowledge, so the learner never recalls or integrates what they know before receiving an answer. Prior Knowledge states the same bypass risk as a general property of generative AI, and adds the design countermeasure — activate before you supply. The fluency illusion is the mechanism on the learner's side: an LLM supplies complete, coherent explanations on demand, so the learner allocates less effort to internal retrieval and reconstruction while the session feels productive, which is precisely the dissociation Tsiligkiris (2026) measured. The three-arm pretesting study makes the same point experimentally: the arm with free AI access and no enforced retrieval performed worst on end-of-semester retention, and interaction volume did not differ significantly across arms, so the gap was produced by structure rather than by how much students engaged.

Preserving retrieval demand. The corpus converges on a small set of design features. The pretesting agent's policy is the most concrete: withhold direct solutions by configuration, probe misconceptions, request elaboration after superficial attempts, and advance only on adequate conceptual engagement. Tsiligkiris recommends embedding retrieval demands after LLM use through closed-tool outputs — short-answer questions, concept maps drawn from memory, teach-back explanations without the model — and separating scaffolding from checking, so the model is available for clarification and Feedback but distinct checkpoints require independent recall. KnowLoop shows the teach-back form in a deployed system: a Peer agent scaffolds reflective teach-back that surfaced gaps learners could not articulate, and its participants asked to move fluidly between clarification and consolidation rather than treating them as strict phases. Chen, Sakhnini and Istead add the assessment-side version — value the construction process rather than the artefact — while Schorr and colleagues (2026) position these tools as learnable skills that reduce overreliance and support learner agency. Underneath all of them is the one non-negotiable: something must require the learner to produce an answer that the model has not already given.

Limits of the Evidence

• Small samples and single contexts. The pretesting study's analytic sample is 89 undergraduates in one applied statistics course, with 27–34 students per arm; the LLM interaction study has n = 22 with a single-group design that supports no causal claim; the teach-back study has 22 participants. The pretesting authors state that replication across domains and task types is needed for generalisability.
• Short or absent retention intervals. The interaction-depth study tested immediate recall only, with no delayed measure, so its null is a null for short-term retention specifically. Memdora's retention advantage and the LLM-depth dissociation alike are not tracked over the weeks that the spacing literature would require; the pretesting study's seven-week window is the longest in the corpus.
• Self-report and proxy measures. Practice effort in the pretesting study is a behavioural indicator scored by two raters against a rubric from conversation logs, which the authors explicitly note is not a measure of internal motivational state; the LLM-depth measure is a keyword-based proxy capturing the surface form of explanation-seeking rather than its quality.
• Instrumentation and modelling limits. Knowledge Tracing models in the corpus fail exactly where spacing decisions need them, and the authors attribute past successes to retroactive full-dataset fitting rather than deployment-realistic validation. The bilingual companion's flashcard quality is unverified against ground truth, with no automated fact-checking layer, and it has run no outcomes study at all.
• Interleaving has essentially no direct evidence here. As stated above, the corpus contains no blocked-versus-interleaved comparison, only a simulation in which Interleaving is one selectable strategy among several.
• Subject populations and content. The evidence base concentrates on higher education, on STEM and statistics material, and on fact-versus-skill distinctions from geometry and multiple-regression content; generalisation to K-12 settings, to humanities and writing, and to non-Western classrooms is not established by these studies.

Connected Concepts

• Desirable Difficulties — the principle these three techniques operationalise
• Cognitive Psychology — memory, encoding and retrieval as the theoretical home of all three
• Prior Knowledge — retrieval practice as activation of what the learner already holds
• Metacognition — judging whether fluency reflects learning
• Self-Regulated Learning — studying consistently what an AI can schedule or answer for you
• Mastery Learning — retention after the mastery threshold, and the spacing that sustains it
• Knowledge Tracing — the models that make or miss scheduling decisions
• Cognitive Offloading — the failure mode when retrieval is delegated to a tool
• Transfer of Learning — the performance–learning gap the techniques address
• Formative Assessment — closed-tool checkpoints as retrieval events
• Productive Failure — attempting before receiving help
• Learners — the stakes of durable rather than fluent learning

Connected Articles

• Do Gains from Generative AI-Enabled Adaptive Pretesting Persist? Evidence from a Retention Study — Akgun & Toker: adaptive spaced retrieval vs fixed retrieval vs learner-directed AI over seven weeks
• Memdora: Designing Cognitively-Grounded Flashcard Interactions for AI-Powered Spaced Repetition — Memdora: FSRS-6 scheduling and a taxonomy of grounded flashcard interactions
• What Students Ask Matters: LLM Interaction Depth, Task Quality, and Immediate Recall in Higher Education — Tsiligkiris: explanation depth predicts task quality but not immediate recall
• Evidence and Theory for why the Best Example-Problem Ratio To Optimize Learning Gain Depends on Knowledge Content — Retrieval practice for facts, worked examples for skills; the content–treatment interaction
• Capturing Session-to-Session Dynamics of Learning and Forgetting: Testing the Limits of Knowledge Tracing Models — Knowledge-tracing models fail to capture the spacing effect under time-based validation
• A Bilingual, LLM-Mediated Lecture Companion for Self-Regulated Learning: Architecture, Theoretical Framework, Comparative and Usability Evaluation, and a Pre-Registered Outcomes Protocol — Automatic flashcard generation removing the authoring cost of spaced practice
• Make or Take: How Students Navigate Self-Created and Instructor-Provided Cheat Sheets — Constructing a cheat sheet as generative study; offloading the construction with AI
• From Confusion to Consolidation: A Staged Conversational Workflow for Post-Lecture Review — Teach-back consolidation surfacing gaps clarification did not
• Simulating Learners' Task-Selection Strategies and System Constraints in Mastery Learning — Interleaving and Blocking as modelled learner strategies under task-selection constraints
• PersonaVLM: Long-Term Personalized Multimodal LLMs — Longitudinal student memory and its open relation to spaced repetition and forgetting curves
• Learning-to-learn in the age of generative AI: A scoping review and conceptual framework — Retrieval practice in the Tools layer of a learning-to-learn framework
• Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning — Pre-fetching agents bypass retrieval practice; the case for intentional friction