Spaced Repetition and the Forgetting Curve

By Brian Tr · Updated Oct 7, 2026

The forgetting curve is probably the most reproduced chart in corporate training. It shows memory dropping steeply after learning, then flattening out. Spaced repetition is the usual remedy: review the material at growing intervals so the drop never gets far. Both ideas are real, and both rest on a century and a half of research.

But the way they show up in slide decks and vendor pages often has little to do with the original work. The famous "you forget 70% within 24 hours" line is a good example. This article goes back to the sources, explains what Ebbinghaus actually measured, summarizes what later research on spacing and testing found, describes how spaced repetition software works, and then turns it into practical guidance for course design.

Key Facts

• Hermann Ebbinghaus published his forgetting data in 1885. He tested himself on series of 13 nonsense syllables and measured memory by "savings", the work saved when relearning, not by how much he could recall (Ebbinghaus, 1885, English translation).
• In his tests, relearning a series after about 24 hours took roughly one third less work than the first learning (33.7%). After six days the saving was about one fourth, and after a month about one fifth (Ebbinghaus, 1885).
• A 2015 replication used one subject, 70 hours of learning and relearning, and intervals from 20 minutes to 31 days. It found results similar to Ebbinghaus's and a possible upward jump in the curve starting at 24 hours (Murre and Dros, 2015).
• A meta-analysis of distributed practice pooled 839 assessments from 317 experiments in 184 articles (Cepeda et al., 2006).
• In one large study of more than 1,350 people, the best gap between two study sessions shrank from about 20 to 40% of a one-week test delay to about 5 to 10% of a one-year delay (Cepeda et al., 2008).
• Students who took a recall test on a passage remembered 56% of it a week later, against 42% for students who restudied it (Roediger and Karpicke, 2006).

What Ebbinghaus actually did

Ebbinghaus was a German psychologist who wanted to study memory the way a physicist studies motion: with controlled material and numbers. The problem with everyday material, such as poems or names, is that people already know parts of it. So he built material with no meaning. He made about 2,300 syllables of the consonant-vowel-consonant type, shuffled them, and drew them at random to form series (Ebbinghaus, 1885, chapter III).

Three details of the design matter for anyone quoting the result.

The material was nonsense syllables. These are meaningless strings. Real course content has meaning, context and links to what a person already knows, and those things change how fast it fades. Ebbinghaus himself noted that the learning of the syllables "falls considerably short" of ideal homogeneity, and that when he learned cantos of Byron's Don Juan the spread of his results looked similar (chapter III).

He was the only subject. The tests ran in two periods, 1879-80 and 1883-84, and he writes of avoiding knowing results in advance and of his own state of mind at the time. The data describe one trained adult, and the author warns that his numbers were affected by time of day and other uncontrolled influences (chapter III, chapter VII).

The measure was savings. He learned a series until he could recite it twice without error, left it alone, then relearned it to the same standard after a set time. The 1879-80 study comprised 163 double tests, each with eight series of 13 syllables, relearned after about a third of an hour, 1 hour, 9 hours, 1 day, 2 days, 6 days or 31 days. The share of the original effort that relearning saved was his measure of what remained (chapter VII).

Savings is not the same as recall. A saving of 33.7% means relearning took about a third less effort than the first time. It does not say that he could recall one third of the syllables or that two thirds had vanished. In the words of the text, after 24 hours "about one third was always remembered" in the sense of work retained, and the forgetting curve was steep early and very slow late (chapter VII).

The curve in his own numbers

The author also flagged that his own 24-hour figure looked "somewhat too large" and might be a point or two lower on a repeat. He called the whole set uncertain and individual (chapter VII). That caution rarely survives the trip to a slide.

Does the curve replicate?

Yes, with qualifications. Murre and Dros (2015) repeated the experiment with a single subject, who spent 70 hours learning and relearning lists of nonsense syllables and was tested after 20 minutes, 1 hour, 9 hours, 1 day, 2 days and 31 days. They concluded that the Ebbinghaus forgetting curve had been replicated, and that it was probably not completely smooth: they found evidence for "a jump upwards starting at the 24 hour data point" that some authors attribute to sleep (Murre and Dros, 2015).

That is useful for two reasons. First, the basic shape is sturdy. It reappears more than a century later, in a different language, with a different person. Second, it is still a single-subject finding on nonsense material. Nobody has shown a single universal curve for all learners and all content. How fast knowledge fades depends on how well it was learned, how meaningful it is, how it connects to other knowledge and whether it gets used.

The 70% in 24 hours claim

You will see versions of this in training decks: "people forget 70% of what they learn within 24 hours" or, for variety, 90% within a week. They are usually attributed to Ebbinghaus.

Compare that with the table above. The 24-hour point in his data is a saving of about one third. A loose reading flips it to "two thirds forgotten", and rounding makes it 70%. But the number is about relearning effort for nonsense syllables, one man and one set of conditions. It is not a measurement of how much of a sales workshop, a compliance course or a software tutorial a typical employee will lose.

The fair statement is narrower: in a classic memory experiment, material learned once and never revisited lost a large share of its benefit within the first day, and the loss then slowed sharply. Later research shows the pattern is real, and that review and retrieval change it. The specific percentage is not a rule for training. If a vendor quotes one, ask what was measured, in whom and with what material.

The spacing effect

The spacing effect is the finding that the same study time produces better retention when it is spread out than when it is packed together. Two pieces of evidence do most of the work.

Cepeda et al., 2006: the meta-analysis

Cepeda and colleagues collected the distributed-practice literature and found 839 assessments of distributed practice in 317 experiments in 184 articles. A central result is that retention depends on the gap between study sessions in a structured way: the interval that produces the most retention increases as the interval over which you need to remember increases (Cepeda et al., 2006). The synthesis leans heavily on verbal recall tasks, so it speaks most clearly to facts and terms.

Cepeda et al., 2008: how big should the gap be?

In a follow-up, more than 1,350 people were taught a set of facts, reviewed after a gap of up to 3.5 months, and tested at a further delay of up to a year. At any test delay, lengthening the gap first improved performance and then gradually reduced it. The best gap grew as the test delay grew, but shrank as a share of it: about 20 to 40% of a one-week test delay, falling to about 5 to 10% of a one-year delay (Cepeda et al., 2008). The authors conclude that many educational practices are highly inefficient.

For a designer, the practical reading is this. There is no single best schedule, because the best gap depends on how long the learner needs the knowledge. A product launch certification needed in two weeks wants short gaps. A safety procedure that matters for a year wants longer ones. And overshooting is costly in a different way: gaps that are too long let learners fall off the curve before the review arrives.

Retrieval practice

Spacing tells you when to revisit. Retrieval practice tells you how. Trying to recall something strengthens memory more than reading it again.

In Roediger and Karpicke's 2006 study, students read prose passages and then either restudied them or took recall tests. Five minutes later, restudying produced better recall. After a week, the tested group recalled 56% of the material against 42% for the restudy group. In a second experiment, students who were tested repeatedly recalled 61% after a week, against 40% for students who simply re-read the passage several times, even though the re-readers felt more confident (Roediger and Karpicke, 2006). Those were university students and short passages in a lab, not employees in a program.

Notice the five-minute result. Retrieval practice doesn't win for immediate performance. It wins for the long run, which is exactly what training is supposed to buy and exactly what end-of-course satisfaction scores can't measure.

How the evidence is rated

A widely cited review by Dunlosky, Rawson, Marsh, Nathan and Willingham (2013, Psychological Science in the Public Interest) rated ten study techniques. Practice testing and distributed practice received the highest utility ratings, while summarization, highlighting and rereading were rated low (press summary of the report). The authors also cautioned that the techniques benefit only students who are motivated and capable of using them, and aren't a cure-all. That is worth repeating: the evidence is strong for these two methods, but it is mostly from education settings and says little about whether adult employees will do optional reviews on their own.

How spaced repetition systems work

A spaced repetition system (SRS) turns the research into a scheduling rule. The learner sees an item, tries to recall it, and rates how well it went. The system decides when to show it again.

The Leitner box. The best-known paper version is credited to the German science writer Sebastian Leitner, who described it in the early 1970s. Cards sit in a series of boxes. A card answered correctly moves to a box that is reviewed less often. A card missed goes back to the first box, which is reviewed most often. No software is needed, and the idea of growing intervals for successes and a reset for failures is the core of most later systems.

SuperMemo and SM-2. Piotr Wozniak, a Polish researcher, wrote the first SuperMemo program in December 1987 and described its algorithm in his 1990 master's thesis, later published online. In the version known as SM-2, the interval after the first repetition is 1 day, after the second 6 days, and after that each interval is the previous one multiplied by an "easiness factor" (E-Factor) for the item. E-Factors start at 2.5 and were not allowed to fall below 1.3, because items that dropped lower were repeated "annoyingly often" and tended to be badly written. After each review the learner grades the response on a 0 to 5 scale, which adjusts the E-Factor (Wozniak, SuperMemo 2 algorithm).

Wozniak describes the formula as heuristic. It's a practical rule built from his own use, not a result proven by a controlled trial. Many modern apps use algorithms that descend from SM-2 or replace it with other models, but the logic is the same: harder items come back sooner, easier items come back later, and the gaps stretch as the item sticks.

What this means for course design

Several design moves follow from the evidence, with the caveat that most of it is about facts and terms.

Plan reviews before launch, not after complaints. If the knowledge has to last, build the review into the course structure: a check after a day or two, another after a week or so, another later. Put the schedule in the design doc next to the learning objectives, as part of instructional design basics.

Make review a question, not a replay. A recap video is re-exposure. A question, a scenario or a flashcard forces retrieval. The research favors the second.

Match the gap to the use. Content needed next week needs short gaps. Content needed at an annual audit needs longer ones. Work backward from when the learner will need to use it.

Don't ask a spaced program to do the job of practice. Spaced retrieval keeps knowledge available. It doesn't build judgment in negotiation or management, which needs practice with feedback. Microlearning is a convenient delivery format for spaced items, but brevity is not the active ingredient. Spacing and retrieval are.

Use rewards carefully. Streaks and points can bring learners back for reviews. They can also push people to tick boxes. See gamification in learning for what to be careful about.

What it means for EdTech companies and L&D buyers

For builders. Spacing needs a scheduler, a bank of questions, feedback and reminders. A catalog of videos with no review layer doesn't get the effect. If you do build it, expect the hard part to be participation: learners must come back on day two and day nine, which is where course completion rates and engagement problems show up. The reminder channel (email, chat, mobile push) matters as much as the algorithm.

For buyers. Ask a vendor to show the review schedule, the recall questions and the data on what share of learners complete the follow-ups. "Uses the forgetting curve" is not evidence. Ask whether retention was tested later, with different questions, and what the comparison group was. The same discipline applies to the outcome claims described in proving training ROI and to how teams evaluate vendors in how companies buy training.

Related reading

• Microlearning
• Instructional Design Basics
• Gamification in Learning
• Course Completion Rates
• Proving Training ROI