The Ebbinghaus forgetting curve:
definition and how to retain information

How Ebbinghaus measured forgetting -- on himself

What makes Ebbinghaus's work remarkable is the rigour of his self-experimentation. He invented nonsense syllables (CVC combinations like "DAX" or "BOK") specifically to eliminate familiarity effects -- he wanted to study pure memory, stripped of meaning and prior association. Over years of experiments, he memorised thousands of these syllable lists, then measured how long it took to relearn them after different intervals.

The "savings score" was his key metric: if a list that originally took 10 minutes to memorise could be relearned in 6 minutes after a delay, the savings score was 40%. This indirect measurement of memory was methodologically elegant -- and his findings have been replicated and confirmed by subsequent generations of researchers using more varied materials.

What the curve actually shows

The forgetting curve is a decreasing exponential curve: it drops steeply in the hours immediately following learning, then flattens progressively. This shape reflects a fundamental property of memory -- recently encoded traces are fragile and lose accessibility rapidly, while memories that survive the initial consolidation window become progressively more stable.

The practical implication is counterintuitive: the timing of the first review matters more than any subsequent review. A review at 24 hours has a dramatically larger impact on long-term retention than the same review at 72 hours. The curve does not wait -- it starts falling immediately after encoding stops.

The concept of memory strength

Later research introduced the concept of memory strength: strongly consolidated information resists forgetting better and requires much longer review intervals before the next review becomes necessary. Poorly consolidated information -- encoded shallowly, without sleep, without review -- decays rapidly and returns to near-zero very quickly.

This concept underlies modern spaced-repetition algorithms: rather than applying a fixed review schedule to everyone, they estimate each individual memory's current strength and schedule the next review to occur precisely when that strength is about to drop below a threshold. The stronger the memory, the longer the interval; the weaker, the sooner it needs reinforcement.

Decay and interference: two distinct causes of forgetting

The Ebbinghaus model describes forgetting as passive temporal decay -- memories simply fade over time without use. But subsequent research revealed a second, often more significant mechanism: interference. Interference occurs when similar memories compete with each other, weakening retrieval of both.

Proactive interference is when older knowledge disrupts the encoding of new material (prior French vocabulary making it harder to learn Spanish). Retroactive interference is the reverse: new learning overwrites or blurs older memories (studying Spanish right after French degrades the French you just encoded). In practice, both types are frequent -- which is why alternating unrelated topics is a better study strategy than massing sessions on similar content.

SM-2: the original formalisation by Wozniak

In the late 1980s, Piotr Wozniak developed the SM-2 algorithm for his SuperMemo software -- the first system to translate the Ebbinghaus curve into a practical, automated review schedule. The insight was simple but powerful: instead of reviewing on a fixed calendar, schedule each item individually based on how well it was recalled last time. Items recalled easily get longer intervals; items recalled with difficulty get shorter ones.

SM-2 assigned each card an "ease factor" that adjusted based on user-rated difficulty, then computed the next review interval using a formula derived from experimental retention data. This algorithm remains the foundation of Anki -- still the most widely used flashcard application in the world -- and has been used by millions of learners across four decades.

FSRS: modelling individual forgetting with machine learning

SM-2, despite its effectiveness, had known limitations: its interval calculations did not adapt to the actual shape of each learner's personal forgetting curve, and it could drift significantly from optimal over long time horizons. FSRS (Free Spaced Repetition Scheduler) was developed to address these gaps. Rather than using fixed formulas, FSRS uses a two-component model of memory -- stability and retrievability -- and refines its parameters through large-scale analysis of real review data.

Stability represents how long a memory can be retained before dropping below a target retention threshold. Retrievability represents the current probability of successful recall given elapsed time since the last review. FSRS optimises interval predictions by fitting these parameters to each learner's actual recall history -- meaning it gets more accurate the more you use it. Studies comparing FSRS to SM-2 show substantially better interval precision and higher long-term retention for equivalent total review time.

1. Review quickly after first learning

The first review should happen quickly--ideally within 24 hours, and certainly before the end of the week. This is where the curve drops fastest, so review has the greatest relative impact. A quick evening pass over notes can cut information loss dramatically.

2. Avoid massed review right before an exam

Cramming creates short-term fluency but very weak retention after 48 hours. Content reviewed all at once the night before follows the forgetting curve from the next day--with a rapid, predictable drop.

3. Accept partial forgetting as normal--and useful

A partially forgotten memory reactivated just before disappearing consolidates more strongly than one reviewed while still very fresh. Partial forgetting is a condition for durable memorization, not an obstacle.

Passive rereading instead of active retrieval

Rereading notes or textbooks creates a feeling of familiarity -- but familiarity is not the same as retrievability. When you reread, you are processing information with external support (the page is right there). When you test yourself, you force your brain to retrieve from memory with no cues. That retrieval effort is what strengthens the memory trace.

The practical difference is significant. Karpicke & Roediger (2008) showed that students who studied a text once then tested themselves repeatedly retained nearly 50% more after a week compared with students who restudied the text four times. Replace rereading sessions with self-quizzing: cover your notes, try to recall, then check. The struggle is the point.

Reviewing too late -- only the night before an exam

Cramming works: you can pass an exam this way. What cramming cannot do is move information into long-term memory. Because it compresses all review into a single massed session, it captures none of the spacing benefit that makes memories durable. The forgetting curve begins dropping from the moment you stop reviewing -- and after a single cramming session, the drop is steep.

Students who cram also tend to suffer from interference: because so much is encoded in a short window without proper consolidation, similar items blur together. The alternative is to spread review across multiple sessions over days or weeks -- even three spaced sessions total outperform five massed sessions in long-term retention.

Ignoring sleep after intensive learning

Memory consolidation is not a passive process that happens automatically during waking hours. It is actively driven by sleep -- particularly slow-wave sleep, during which the hippocampus replays newly encoded memories and transfers them to cortical long-term storage. Learning a complex topic and then staying up late, or sleeping fewer than 6 hours, can undo a significant portion of the day's encoding.

The practical implication: if you have an important learning session in the evening, prioritise sleep that night rather than squeezing in more review. A well-slept brain consolidates more effectively than an exhausted one reviewing for an extra hour.