Spaced Interleaved Retrieval

Spaced Interleaved Retrieval is the practice of producing knowledge from memory in order to find out which parts of it were not retained. Retrieval means producing the answer, spoken, written, or solved, before the source is opened. Spacing leaves days between those sessions. Interleaving comes at the same knowledge from a different angle on each session.

The sessions have two jobs. The first job is finding gaps early. A gap costs more the longer it stays unfound. Between reading something and having to produce it, a misunderstanding gives no signal. Everything learned on top of it inherits the error. The time left for repairs keeps shrinking. The second job is building fluency to the standard the use demands. Fluency means how fast the knowledge comes and how complex a use it can support. At the low end, simple recall takes half a minute of searching. At the high end, the knowledge comes almost instantly and can be used inside a complex problem. Both jobs run for the whole life of the material. The sessions are not something that starts a few weeks before an exam.

Core takeaways

• Every session runs the same five moves: produce the material in full without looking anything up, mark every spot that felt shaky, check the attempt against a source or key, decide for each miss between another retrieval pass and re-encoding, and let what failed set the next session’s agenda.
• Most of the measured value is in the checking. Testing with feedback measures about d = 0.73; without feedback it drops to d = 0.39. An unchecked session is worth about half a session, so the answer key is built before the session is scheduled.
• Interleaving means changing the angle between passes, not only mixing subjects. Mixed practice helps at g = 0.42 overall, does nothing measurable on expository text, and reverses to about g = −0.39 on word learning, so near-identical items are blocked and problem types are interleaved.
• Session design sets three dials: order, volume, and type. Over a topic’s life the dials follow a funnel: high-order and high-volume early, mixed-order and low-volume late.
• The layers of the schedule matter more than their dates. A schedule that is off by a few days still works; a schedule with a missing layer does not. The gaps widen as the material ages: a first pass within a day or two, another a week to ten days later, another three to four weeks after that.
• Retrieval is half of a two-part system. Encoding is the permanent fix and takes months to years to build. Retrieval keeps performance up while the encoding improves. Retrieval time is never cut on purpose; it falls on its own as the encoding improves.

One session

The single session is the unit the schedule is built from, and every session runs the same five moves. First, the material is produced before anything is looked up, and it is produced in full: written, spoken, drawn, or performed, never answered silently in the head. A silent mental answer feels sufficient because familiarity feels like capability. It reliably fails to find gaps, and it uses up practice questions that then cannot be reused. Second, every spot where the attempt felt shaky is marked, even when the final answer was right. An answer that came out once, with uncertainty, will fail when the question is varied. Third, the attempt is checked against the source, an answer key, or a worked example. Fourth, each miss gets a decision: another retrieval pass, or a return to how the material was encoded. Fifth, what failed sets the next session’s agenda.

A gap is anything that could not be produced, came out wrong, or felt unsure, including what eventually came out right.

Most of the measured value is in the checking. In the meta-analytic record, testing with feedback measures around d = 0.73. Without feedback it drops to d = 0.39. An unchecked session is therefore worth roughly half a session, so the answer key is built before the session is scheduled. A failed attempt followed by feedback beats never attempting. The attempt is what makes the feedback useful. At least one method in rotation checks against something outside the learner’s own head: an official key, a question bank, or a model answer built against the real assessment scope. Self-testing can miss the same gaps for several rounds, because the learner’s sense of what the scope includes is part of what is missing. Recognition-style checking most often misses transfer. Recall of the practised form usually holds while use beyond that form fails. Only full production shows that difference.

Deliberate retrieval also has a small form, called micro-retrieval: recalling or explaining something within seconds or minutes of first meeting it, inside the learning session. In practice it takes the shape of note-taking. Think the material through first. Write only once it makes sense. Keep the notes as keywords, arrows, and sketches rather than sentences.

Changing the angle between passes

Interleaving means more than mixing subjects. The angle changes between passes. The same material can be taught on one pass, used to solve a problem on the next, and argued against a rival explanation on a third. That is already interleaving. The knowledge is one thing that can be approached in several ways, and the sessions take the ways in turn.

Mixing helps only within a specific range: two things similar enough to be confused and distinct enough to be told apart. Working both from several angles shows where the line between them is. Much of the gain is learning what a concept is not. Mixing fails at both ends of the range. Where most people would call the two things one thing, there is nothing to separate. Where the two things were never confusable, there is nothing to learn. The outside record shows the same range. Mixed practice helps at g = 0.42 overall. It does nothing measurable on expository text. On word learning it reverses to about g = −0.39, and blocking wins. So near-identical items, vocabulary above all, are blocked, and problem types are interleaved.

Three dials and the funnel

Designing a session means setting three dials. Order is whether answering requires two or more concepts and the relationship between them. A problem in context is high-order; one isolated fact is low-order. Volume is how much ground the method covers. A whole topic taught is high-volume; two concepts explained is low-volume. Type is declarative or procedural, matched to the subject’s real balance. Declarative, Procedural, and Conditional Knowledge carries that distinction in full, and Interleaving - Multiple Angles and Session Design works all three dials at length. The two highest levels of the order dial are identifying relationships, then judging which of them matter most. Judging which relationships matter most is the level that sessions most often fail to test. The full set of levels is at Knowledge Mastery - From Recognition to Usable Knowledge.

Across a topic’s life the dials follow a funnel. Early sessions are high-order and high-volume. Late sessions are mixed-order and low-volume. Type is matched to the subject throughout. The reason for the funnel is the cost of finding one gap. Two days after learning, teaching the whole topic out loud produces a gap every few sentences. Months later, the same full dump might produce one gap per page of output, and targeted questions aimed at known weak spots find gaps at lower cost. During the early wide passes, low-order gaps are noted and passed over rather than filled on the spot. The list of noted gaps becomes the next session’s agenda, and that is where the funnel narrows. An integrated, high-order foundation is retained and gives later detail a place to attach. Detail learned first stays a set of unconnected items. Higher-order methods find both kinds of gap. Lower-order methods find only low-order gaps.

Declarative or procedural

The type dial is where self-diagnosis goes wrong most often. Applying knowledge to a case is still declarative. Procedural means execution that has become automatic: driving without thinking about the pedals, speaking without assembling the sentence, writing code without looking up the syntax. Most subjects contain less procedural knowledge than they appear to. Even in subjects everyone calls procedural, the usual bottleneck is declarative. Isolated problems are solved cleanly, then a problem in context stalls on which tools apply and in what order. That is a conceptual gap that looks like a practice-problem gap. The working diagnostic is one complex challenge, classified by how it was missed. A wrong approach, a wrong strategy, a wrong ordering, or no understanding of how the ideas connect means the gap is declarative. The right approach with wrong arithmetic, sound architecture with a syntax bug, or strong ideas in clumsy sentences means the gap is procedural. Where the bottleneck really is procedural, varied execution starts early. Mixed problem sets beat blocked ones 61% to 37% on a test a month later, around d = 0.83. Interleaving for Complex Problem Solving treats that mixing as reconstruction practice rather than as scheduling.

Picking the method

Methods are chosen from the dial settings, not by preference, and no subject gets every method. Interleaving Table holds the full catalog. The working core of that catalog:

What is being tested	Session forms	Watch for
Isolated facts, labels, constants	Narrow flashcards, image occlusion, quick written recall; cover-copy-check: cover the notes, reproduce them from memory, uncover and compare	Too low-order for complex understanding. The public evidence for cover-copy-check sits in spelling and arithmetic fluency, which is exactly as narrow as this row
One concept	Teach it aloud without notes, then check; single-concept generated questions	Stays low-order if the questions stay narrow
Relationships	Relational questions: compare, contrast, connect; a local map rebuilt from memory	Needs full answers. “I could have answered that” is familiarity, not evidence
The big picture	Closed-book brain dump: everything recallable written out with the source shut, as running text or a branching map; a chunkmap rebuilt from memory, meaning a map of a few named clusters and the cross-links that matter most	Linear dumps under-represent relationships. Maps can miss fine detail. A hard rebuild means the encoding was shallow
Judgment	Evaluative questions on importance, relevance, and priority; directed discussion arguing which relationship matters more	Time-costly, so the questions get written in one session and answered in another. Both halves are retrieval, and the first sets up the second
A procedure	Varied problems, execution under realistic constraints, simulation of the target task, or working backward from a target output. Change the values first, then add conditions that change the structure: the same problem re-asked with a new condition each pass	Repetition past first success gains nothing. Vary the practice or stop it
Exam readiness	Timed mixed questions, full answer generation, error review	Checking too fast records a real gap as a silly mistake and leaves its cause intact

The teaching-shaped methods are one act at four volumes. The smallest is explaining a thing simply until the point where the explanation stumbles. That is a repair tool, not a system. The next two are explaining one concept, then explaining how several concepts interact. The largest is WPW, whole-part-whole: the big picture from memory, each part in detail, then the big picture again with the parts interacting. WPW belongs early in a topic’s life, where the funnel is wide. Teaching is done to an imaginary student, never to a friend. A friend fills the gaps, and finding the gaps is the reason for the method.

The extended practice question and its group form

The strongest single method among these is the extended practice question. It has four moves: answer from memory, mark every point of low confidence, build your own best answer from the material, then check that answer sheet against the official one. Elaborated production of exactly this kind is one of the three factors that decide whether retrieval transfers at all. The group form uses the same steps with several people. Each person writes a hard exam with model answers. The exams are exchanged and answered from memory. Then the answer sheets are exchanged and the differences are argued. Its consolidation per hour is among the highest of any method here. Undirected group work does worse than that structure. Small-group work in general still measures well, around d = 0.51 in the STEM record, so the structure is what is kept. Memory palaces and related mnemonics stay narrow by design. Trained mnemonic work can roughly double recall on a word-list task, and the gain is confined to that kind of demand: ordered lists, labels, rote sequences. They do not work as a substitute for encoding.

Overlearning and the methods to leave out

The system’s name for the main failure mode is REBIM, Repetitive Execution Beyond Initial Mastery. The public literature calls it overlearning. It means continuing to practise a procedure that is already performed smoothly, without changing its difficulty, constraints, or context: coding the same simple scripts, solving problems at a difficulty that is already comfortable, running the same conversation pattern in a language that already handles it. Its return has been measured. Six extra problems past the first success produced no score gain at one week or at four weeks. The same practice moved a week later produced a large gain at four weeks. The reason people keep doing it has also been measured. Learners rated blocked practice as more effective immediately after a test on which they had done better under interleaving. From the inside, smooth performance cannot be told apart from progress. The excluded methods fail for the same reason. Passive rereading rates as low utility in the technique-comparison literature. Rewriting and relistening work the same way as rereading. Mental-only answers, and answers checked before a full attempt, fail for the reasons given above. The Shortcut Problem explains why the version of each method that feels cheaper is the one that fails.

Where an AI model fits

An AI model can take the low-cognitive work: question variations, mixed sets, counterexamples, completeness checks, follow-ups after a weak answer, prompts converted from a map, questions graded by difficulty, keyword collection, and big-picture orientation. It can also consolidate several sources into one. That removes the split-attention cost of working across documents, which is one of the most replicated results in cognitive-load research. Guardrails decide whether the model helps or harms. A tutor configuration that would not hand over answers raised practice performance 127% with no later harm. Unrestricted access raised practice scores 48% and cut unaided exam performance 17%, compared with students who never had it. With unrestricted access the model did the structural thinking, so the learner never built the schema. Beginners are the most exposed, because they have not yet built a reliable sense of their own cognitive load. Checking the truth of nuanced claims is outside a model’s reliable range for now. The rule that keeps the benefit is: produce first, in full, then give the answer to the model for feedback and calibration. When the model does the thinking and the learner only edits, the benefit of every method above disappears.

Retrieval and encoding

None of these sessions works alone. Retrieval is half of a two-part system. In the research, retrieval and encoding are one process; in practice they are split only because the techniques differ. Retrieval is the half that gets trained first. Encoding is how material is structured while it is being learned. Encoding is the permanent fix for forgetting, and it is a skill that takes months to years to build. Retrieval is repetitive by nature, usable within a day or two of learning it, and it keeps performance up while the encoding skill is being built. That is why ICS System trains retrieval first. The system’s operating defaults give the stakes. Without retrieval, around half of new material is gone within a week. Retrieval takes upward of 70% of study time for most people starting out, under 20% for a practised encoder, and never zero. Weaker encoding increases the retrieval load: more decay, more frequent and longer sessions, and roughly five times more repetition per gap, on the system’s own estimate.

The sessions also measure the encoding. A retrieval failure shows whether the Bear Hunter System pass produced a structure that holds. That pass sets questions before the source is read, builds the answers into a relational map, and cleans the map until it can be rebuilt from memory. The questions written in that pass become the prompts for the retrieval session, and the finished map serves as the answer key. Many low-order gaps at once point to a problem with that structure before they point to a need for more cards. Better chunking removes isolated memorisation at the source. The card-formulation practitioners arrived at the same rule independently. Early sessions run long and find many gaps, which is the expected state. Many gaps found means the session is working. Few gaps found while mastery is still low means the probe is too shallow. The sessions shorten on their own as the encoding improves. Encoding and Retrieval holds the tradeoff in full.

The schedule

The layers of the schedule matter more than their dates. A schedule that is off by a few days still works. A schedule with a missing layer does not. That is the system’s own rule: session quality matters more than scheduling precision. The outside record sets limits on it. The useful gap between sessions depends on how long the knowledge has to last: roughly 20–40% of a one-week retention interval, falling toward 5–10% of a one-year one. So the interval is set from the date the knowledge is needed. Each mistiming has a specific cost. Sessions that are too frequent produce repetition that taught nothing, because the test came before forgetting began. Sessions that are too infrequent turn into relearning, and into sessions the learner does not want to start.

The gaps widen as the material ages: a first pass within a day or two, another a week to ten days later, another three to four weeks after that. The reason for the widening is workload. With new material arriving daily, a fixed interval builds a backlog without limit. A longer gap is also a harder and more realistic test. Knowledge is normally used long after it was learned, and forgetting a little between passes is what makes gaps visible. The widening itself has no retention evidence behind it. Equally spaced retrieval has beaten expanding schedules at a two-day test and matched them on text. So the widening is a logistics decision, and it is made knowing that.

Inside a study week the layers fall in this order. Before new material: prestudy, a fast and shallow pass over the concepts and how they relate, with nothing memorised. On the learning day: a light same-day pass. This pass supports encoding more than it serves as a memory-optimal first test, because the retention evidence favours delaying the first real test. The day’s material is revised and its detail built. First cards are started for the isolated facts. Definitions are left alone for now. Within a day or two: the first gap-finding pass. End of the week: the last two weeks’ material by an interleaved method, weak spots revisited, and detail memorisation starting here. End of the month: the four weeks again in a different form, which by then should feel easy. Two to four weeks before an assessment: a comprehensive pass over everything relevant. The system’s own guidance gives that two-to-four-week range. Heavy content adds a mid-week or fortnightly layer. Cards run in spare minutes during the day rather than inside this schedule, because the card apps stage the widening themselves. Hours spent on cards is a sign that the card count is too high, which points back at the encoding. Flashcards covers that layer on its own. Multipass System runs the same sequence in compressed form. Cramming is what this schedule turns into when the time before the assessment gets short.

Counted across the layers, any piece of material is produced or re-encoded about six times in a month: prestudy, the learning event, the same-day pass, the mid-week pass, the weekly pass, and the monthly pass. That six is a count of the schedule’s layers, not a measurement. The time budget depends on one premise: at 80–90% retention by the weekly pass, a week’s material takes a few hours and a month’s takes about half a day. Below that retention, sessions run longer. The extra length is information about the encoding, not a personal failing. The schedule also stops where the use stops. For material needed once, good enough is the correct target. The fluency bar always comes from the reason the material is being learned.

Repairing each kind of gap

Every session ends with a gap list. Each entry gets a repair matched to its type.

The gap	The repair
A fact would not come	A narrow card, or a quick written recall check
A relationship would not come	Rebuild the local map from memory and compare
The structure would not come	Rebuild the topic’s whole structure from memory, compare it against the notes, then rewrite the notes into that structure. The organisation failed, so the organisation gets repaired
Known material failed in a new shape	Raise accuracy on the base form first, then elaborate the production (explain, justify, extend), and only then vary the format
The answer came, without confidence	Build a full answer sheet and compare it against the official one
The procedure stumbled	Practise the execution itself, with feedback

The row for known material failing in a new shape follows directly from the evidence. Retrieval transfers at about d = 0.40 overall. The transfer model predicts no gain at all when three things are absent together: congruent response demands, elaborated production, and high initial accuracy. So adding more interleaving to a foundation that is not yet accurate does not fix transfer. Practising edge cases on their own does not fix it either, although edge cases look like the transfer fix.

The first repair for any gap is the same method again. A failed test on its own makes the material easier to retain afterward. The second pass is more focused because the miss is known. Often, going back over the material knowing how it went wrong is the whole repair. A repeat failure means the connection itself was weak or wrong. The next step is then to build a different connection on purpose: step slightly outside the scope, tie the item to problems already understood, or force the wide view with an analogy. Some items take several rounds of that. Several rounds is the normal course, not a sign to repeat the same route harder.

A session also needs a stopping rule, and the relearning literature supplies one. Recall each item until it has been produced correctly three times. Then relearn to the same criterion across about three widely spaced sessions. The first relearning session takes under two minutes per concept, and later sessions take less. That gives a per-item budget instead of an ending set by tiredness.

Some sessions cost no extra time, because the working day supplies them. Any real use of the knowledge is a retrieval session: building the plan, judging an approach, taking the wide view. What makes such a use count is labelling it afterwards by order and type. A day that only demanded facts leaves the higher-order side untested, and a planned session then covers that. Where the order of study is free but the order of work is fixed, studying what the coming weeks will demand collects those repetitions at no extra cost. Opportunistic Retrieval covers that practice.

Where the system reaches its limit

The main case against the system is its limit. Retrieval is only as good as the encoding under it. Weak encoding means many gaps, long sessions, and the five-fold repetition per gap accumulating over time. That accumulation is where the feeling of relearning the same material forever comes from. A study system whose only real technique is testing reaches a limit that more work cannot get past, because the work grows faster than the time available. The way past the limit is Deep Processing, the dimension that decides how much retrieval is needed at all. Deep Processing Practice is the training ground that cuts the gap count at its source. The costs are concrete even when the limit is far away. The layers cost the hours given in the schedule above, and they hold to those hours only on the stated retention premise. A session worth running needs its key or answer sheet built first. The group protocol adds coordination on top.

Signs it has gone wrong and signs it is working

The signs that the sessions have gone wrong are specific. When checking the notes starts to feel easier than producing the answer, the sessions have become recognition. The tendency toward that state has been measured: learners reliably prefer the condition that teaches them less. Are You Thinking, or Just Consuming is the standing test for it. Few gaps found while mastery is still low means the probe is too shallow, never that the material is retained. Revision carries the session shapes that keep a probe accurate. Sessions that have become smooth repetition of what already works are REBIM on a schedule.

There are checkable signs that the system is working. Recall feels effortful but reachable: controlled strain, not guessing at random. Gaps stop being vague and become specific. Similar ideas get easier to tell apart as the mixing separates them, and the earlier limit still applies: near-identical material stays blocked. Spacing shows which knowledge is actually stable rather than only recently seen. Every failure points to a next step, either another pass or re-encoding. One more thing to expect is a drop in confidence. When it becomes clear why re-reading was producing nothing, confidence drops before it rises, because the problem is now visible. The drop is expected and it passes. The system’s own practice plan allows for that period: two to four weeks of running several kinds of session before any further technique is added. The skill is mostly selection and honesty, and both come from repetition. Technique added on top of unpractised retrieval produces a collection of methods that nobody runs.

The number to watch

Retrieval time is never cut on purpose. It falls on its own. Run this way, the sessions produce knowledge that can be rebuilt and used under conditions it was never practised in. They also produce one number worth watching: how much retrieval the material still needs. As encoding improves, sessions find fewer gaps and shorten by themselves. Deciding to do less retrieval because the encoding should be better by now lets forgetting return without being noticed. The retrieval time still being spent is the measurement of what the encoding is failing to hold. Total study time may stay the same for weeks or months while that measurement falls. That is the normal course. Accuracy comes first. Efficiency follows from accuracy, not the other way around.

How to practice this

1. Produce the material in full before opening any source: write it, say it, draw it, or solve it. Notice where the answer felt shaky, even where it came out right. Mark each of those spots as a gap.
2. Check the attempt against the source, an answer key, or a worked example built before the session. Notice that a session without this check is worth about half a session.
3. For each miss, decide between another retrieval pass and a return to how the material was encoded. Try the same method again first. Notice whether the item fails a second time, which means the connection was weak or wrong.
4. On the next pass, change the angle: teach it, solve with it, or argue it against a rival explanation. Notice which similar ideas get easier to tell apart. Keep near-identical items such as vocabulary blocked.
5. Space the passes: within a day or two, a week to ten days later, then three to four weeks after that. Notice whether recall feels effortful but reachable. Notice whether the sessions shorten on their own as the encoding improves.
6. Stop each item after three correct productions, then relearn it to the same criterion in about three widely spaced sessions. Notice that the first relearning session takes under two minutes per concept and later ones take less.

Related pages

• Retrieval: the dimension these sessions sit inside.
• Self-Regulation: where the decision to run, keep, or change a session gets made.
• Metacognition - The Control Layer: judging what a failed session means before the plan changes.
• Dimensions of Learning: the map both dimensions belong to.
• Prestudy, BHS, and SIR - Turning Information into Usable Structure: the whole pipeline walked end to end.

3 more related pages

• The Technique Is Only as Good as the Thinking It Produces: why a correctly performed session can still produce nothing.
• Reconstruction - Retrieval Beyond Recall: why producing the structure beats producing the item.
• Memory Handling: what memory does with what the sessions hand it.

Open questions

• How is a procedural bottleneck told from a declarative one before the first complex challenge exists to fail at?
• Where does the line fall between a gap that wants another retrieval pass and a gap that wants re-encoding?

Sources

• Rowland (2014), Psychological Bulletin. Testing-effect meta-analysis: recall beats recognition; with feedback d = 0.73, without d = 0.39.
• Cepeda, Vul, Rohrer, Wixted & Pashler (2008), Psychological Science. The spacing ridgeline: optimal gaps near 20–40% of a one-week retention interval, 5–10% of a one-year one.
• Karpicke & Roediger (2007, JEP:LMC; 2010, Memory & Cognition). Expanding versus equal spacing: equal wins at two days, no difference on text; delaying the first test helps long-term retention.
• Latimier, Peyre & Ramus (2021), Educational Psychology Review. Meta-analytic support for spacing out retrieval episodes.
• Rohrer & Taylor (2006), Applied Cognitive Psychology. Overlearning null at one and four weeks; the same practice distributed buys a large delayed gain.
• Rohrer, Dedrick, Hartwig & Cheung (2020), Journal of Educational Psychology. Preregistered classroom RCT: interleaved 61% vs blocked 37% one month later, d = 0.83.
• Brunmair & Richter (2019), Psychological Bulletin. Interleaving meta-analysis: g = 0.42 overall, −0.39 on word learning, similarity as the governing moderator.
• Kornell & Bjork (2008), Psychological Science; Birnbaum, Kornell, Bjork & Bjork (2013), Memory & Cognition. The fluency illusion and the discrimination account.
• Kornell, Hays & Bjork (2009), JEP:LMC. Failed retrieval plus feedback beats never attempting.
• Pan & Rickard (2018), Psychological Bulletin. Transfer of test-enhanced learning: d = 0.40 overall; response congruency, elaborated retrieval, and initial accuracy decide it.
• Rawson & Dunlosky (2022), Current Directions in Psychological Science. Successive relearning: the three-correct-recalls criterion and its time costs.
• Dunlosky, Rawson, Marsh, Nathan & Willingham (2013), PSPI. Rereading rated low utility.
• Fiorella & Mayer (2013), Contemporary Educational Psychology; Bisra, Liu, Nesbit, Salimi & Winne (2018), Educational Psychology Review. Learning by teaching and self-explanation.
• Springer, Stanne & Donovan (1999), Review of Educational Research. Structured small-group learning in STEM, d ≈ 0.51.
• Dresler et al. (2017), Neuron. Mnemonic training roughly doubles word-list recall, confined to that demand.
• Ayres & Sweller, in the Cambridge Handbook of Multimedia Learning. The split-attention effect.
• Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman (2025), PNAS. Guardrailed tutoring +127% practice with no later harm; unrestricted access +48% practice, −17% unaided exam.
• Woźniak, Twenty rules of formulating knowledge. Card bloat as a formulation and comprehension failure.
• Tulving & Thomson (1973), Psychological Review. Encoding specificity; Chase & Simon (1973), chunking capacity.
• The system itself, meaning the session moves, the dials and funnel, the schedule and its layers, and every operating default on this page (the ~6 passes a month, the 80–90% retention premise, the ~5× repetition per gap, the 70%-to-20% share of study time, the half-gone-in-a-week figure), is the learning doctrine this wiki organises, stated as its own.