Spaced repetition (distributed practice) means returning to the same material several times with gaps in between, rather than studying it all at once, in order to retain it for longer.
Instead of reading something ten times in a row, you come back to it on day one, day three, a week later, a month later. The gap itself is the mechanism — it changes how much survives.
After a gap you have half-forgotten the item, and that difficulty in retrieving it is what strengthens the memory. Cramming looks better on a test taken immediately and worse on one taken weeks later; spacing draws the opposite curve.
In practice this needs a scheduler that decides when each card comes back. Anki's SM-2 and the newer FSRS are the well-known ones.
Only peer-reviewed work is cited here.
Ebbinghaus memorised nonsense syllables using himself as the sole subject and measured how much he retained over time. That is the forgetting curve, and it is where the idea starts.
It is also a self-experiment with one participant on nonsense syllables. Nothing guarantees the numbers transfer to learning real vocabulary.
Cepeda et al. (2006) synthesised the distributed-practice literature for verbal recall and report that spaced study beats massed study on retention. Being a review and quantitative synthesis rather than a single experiment, it is stronger evidence than any one study.
Cepeda et al. (2008) showed that the optimal gap is not a constant: it scales with how far away the test is. The longer you need to remember something, the longer the gaps should be.
🔴 This is the practically important one. There is no single correct interval — which is exactly why a scheduler exists.
Settles and Meeder (2016) trained a retention model (HLR) on Duolingo's large-scale review logs, treating interval selection as a machine-learning problem. It is one of the reference points for scheduler work since.
Claims made in the name of spaced repetition that the evidence does not support.
What Cepeda et al. (2008) showed is that the optimal gap depends on the retention interval you care about. A table you found somewhere has no particular reason to match your goal.
Ebbinghaus ran on one subject with nonsense syllables. Carrying "you forget 42% in 20 minutes" straight over to foreign-language vocabulary is a leap.
What is established is long-term retention. It is not a finding that total study time goes down. For an exam next week, cramming can genuinely win.
"The predicted review time was right" and "this person learned more" are not the same quantity. What we publish is the first one.
LangDict uses spaced repetition, and we check that decision by measuring prediction accuracy.
If the optimal gap depends on the situation, a fixed table cannot be the answer — a model has to decide. In LangDict, our own model Lexa sets the next review date.
Lexa reads not only the card being predicted but the review history of that user's other cards as context, which lets one shared model personalize without a per-user fitting phase.
We publish the comparison against the default settings of FSRS-7 on three public benchmarks (Anki / Duolingo / maimemo), together with the metric, the split and the scores. What is being compared is retention-prediction accuracy, not learning outcomes.
See LangDict →The Lexa research report →Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Leipzig: Duncker & Humblot.
Only works whose author, year and venue we have checked against the original are listed. For journal articles, volume, issue and pages were checked against Crossref as well.
Last updated: July 2026