The science

What the research says, and what it doesn't.

Dyscendi rests on three findings about memory and one scheduling algorithm. This page takes them one at a time: what was tested, on whom, what came out, and where the result stops. Every figure links to its source.

Forgetting
Spacing
Testing yourself
Study techniques compared
FSRS
What this doesn't prove
References

1 · Forgetting

Memory fades on a curve, fast at first.

In the 1880s Hermann Ebbinghaus learned lists of nonsense syllables, waited, and measured how much quicker he could relearn them. That saving is the trace the first learning left behind.

The forgetting curve, replicated 130 years later

Murre & Dros, 2015 · PLOS ONE^[1]

Question: Does Ebbinghaus's curve hold up when someone else repeats his experiment?
Method: One subject spent 70 hours learning lists of nonsense syllables and relearning them after 20 minutes, 1 hour, 9 hours, 1 day, 2 days, 6 days or 31 days.
Result: The two curves nearly overlap. After one day, the saving was 22.3% for Ebbinghaus and 21.6% in the replication. After 31 days, 5.0% and 4.1%.
Limit: One subject, like the original, and nonsense syllables rather than course material. The authors also note the curve isn't perfectly smooth: it shows a jump upwards around the 24-hour point.

Saving on relearning, by delayShare of the original learning effort saved when relearning the same list. Higher means more was retained. Table 3 of the paper.
Delay	Ebbinghaus (1880s)	Replication (2015)
20 minutes	58.6%	61.2%
1 hour	47.6%	47.9%
9 hours	31.7%	27.2%
1 day	22.3%	21.6%
2 days	18.0%	19.8%
6 days	9.6%	9.4%
31 days	5.0%	4.1%

What Dyscendi takes from it: a review is worth most just before the memory slips, and that moment arrives quickly for something new.

2 · Spacing

The same reviews work better spread out.

Cramming and spacing can cost the same number of minutes. What changes is how long the result lasts.

A review of 317 experiments

Cepeda, Pashler, Vul, Wixted & Rohrer, 2006 · Psychological Bulletin^[2]

Question: Across the published research, how does the time between study sessions affect what people remember later?
Method: A meta-analysis: 839 assessments of distributed practice, from 317 experiments in 184 articles.
Result: The best gap depends on how long you need to remember. In the authors' words, the gap “producing maximal retention increased as retention interval increased.”
Limit: The review covers verbal recall tasks — words, facts, texts. It doesn't speak for motor skills or problem-solving.

How long should the gap be?

Cepeda, Vul, Rohrer, Wixted & Pashler, 2008 · Psychological Science^[3]

Question: If you get one review, when should it fall to remember the material at a given later date?
Method: More than 1,350 people were taught a set of facts, reviewed them after a gap of up to 3.5 months, and took a final test up to a year later.
Result: There is a best gap, and it stretches with the delay. Measured as a share of the delay before the test, it went from about 20 to 40% of a one-week delay to about 5 to 10% of a one-year delay.
Limit: One review, one kind of material. A real course means many reviews of many cards, which is the problem a scheduler exists to solve.

What Dyscendi takes from it: intervals that grow as a card gets stronger, instead of a fixed “every three days”.

3 · Testing yourself

Pulling an answer out of memory strengthens it more than reading it again.

This is why a flashcard hides its answer. The effort of recalling is the exercise.

Re-reading feels better. Testing works better.

Roediger & Karpicke, 2006 · Psychological Science^[4]

Question: Is it better to read a text again, or to try to recall it?
Method: In the second experiment, 180 undergraduates studied a prose passage four times, or three times plus one recall test, or once plus three recall tests. No feedback was given on the tests.
Result: Five minutes later, re-reading won. A week later, it lost badly. The numbers are in the table below. The re-readers were also the most confident that they would remember.
Limit: Short texts, a one-week horizon, university students in a lab.

Share of the passage recalledExperiment 2, 30 students per group and delay.
How they studied	After 5 minutes	After 1 week	Forgotten in a week
Read four times	83%	40%	52%
Read three times, one test	78%	56%	28%
Read once, three tests	71%	61%	14%

Once you know it, keep testing it

Karpicke & Roediger, 2008 · Science^[5]

Question: After you've recalled something correctly once, is it more useful to keep studying it or to keep testing it?
Method: College students learned 40 Swahili–English word pairs. Once a pair was recalled, it was either kept in the tests or dropped from them, and either kept in the study rounds or dropped from them.
Result: Testing made the difference; extra studying did not. A week later, students who kept being tested recalled about 80% of the pairs. Those whose learned pairs were dropped from testing recalled 36% and 33% — even the group that went on studying every pair.
Limit: Vocabulary pairs, one week. The students also failed to predict any of this: their forecasts were the same in every group.

What Dyscendi takes from it: a card you got right comes back anyway, later. Getting it right once isn't the finish line.

4 · Study techniques compared

Of ten common techniques, two came out on top.

The techniques students use, ranked by evidence

Dunlosky, Rawson, Marsh, Nathan & Willingham, 2013 · Psychological Science in the Public Interest^[6]

Question: Which study techniques actually help, across ages, subjects and kinds of test?
Method: A review of the research on ten techniques, each rated for overall usefulness.
Result: Practice testing and distributed practice received the highest rating. Five techniques were rated low, among them three of the most popular: summarising, highlighting or underlining, and re-reading.
Limit: A rating sums up many studies. It can't say what will work for one student in one course.

A flashcard reviewed on a schedule is both top-rated techniques at once: a self-test, spread over time.

5 · FSRS

The scheduler that picks the day.

Knowing that spacing works doesn't tell you which day to review which card. That's a prediction problem: how likely are you to remember this card tomorrow, or in three weeks? FSRS — the Free Spaced Repetition Scheduler — is an open-source answer to it, and the one Dyscendi uses.

Where it comes from

Ye, Su & Cao, 2022 · ACM SIGKDD Conference on Knowledge Discovery and Data Mining^[7]

Question: Can a model of long-term memory, fitted to real review data, schedule reviews at a lower cost?
Method: A memory model built from 220 million logs of students' memory behaviour, and a scheduler designed to minimise the review cost. The scheduler was deployed in the language-learning app MaiMemo.
Result: A 12.6% improvement over the methods it was compared with, as reported by the authors. The FSRS project describes its own research as following the path this work opened; the algorithm has been revised several times since.
Limit: The comparison is the authors' own, in the setting of one language-learning app.

How it measures up today

Open Spaced Repetition, SRS Benchmark · consulted 12 October 2026^[8]

Question: Given someone's past reviews, how well does each algorithm predict whether they will remember a card at the next one?
Method: A public benchmark on 9,999 collections, each from one Anki user, and 349,923,850 reviews. Data is split by time: older reviews train the algorithm, newer ones test it.
Result: FSRS-6 predicts recall better than the older and simpler schedulers — in 98.3% of collections against half-life regression, the model Duolingo published. It is not the best in the table: FSRS-7 and several neural networks do better.
Limit: See the notes under the table. They matter.

Predicting recall: FSRS-6 and a selection of other algorithmsLower error is better. The last column is the share of collections in which FSRS-6 had the lower log loss. A selection of rows; the full table is in the benchmark.
Algorithm	What it is	Log loss	RMSE (bins)	FSRS-6 more accurate in
LSTM	A recurrent neural network	0.3332	0.0538	11.9%
FSRS-7 (recency)	The newest FSRS version	0.3370	0.0593	17.3%
FSRS-6	The version Dyscendi uses	0.3460	0.0653	—
FSRS-5	The previous version	0.3561	0.0742	84.2%
FSRS-4.5	An earlier version	0.3625	0.0764	92.4%
DASH	A model from academic research	0.3682	0.0838	89.0%
ACT-R	A cognitive-architecture model of memory	0.4033	0.1074	95.3%
HLR	Half-life regression, proposed by Duolingo	0.4694	0.1275	98.3%
Ebisu v2	A Bayesian half-life estimator	0.4989	0.1627	97.8%

Reading the columns. Log loss measures the gap between the predicted chance of recall and what actually happened at each review. RMSE (bins) groups reviews by interval, number of reviews and number of lapses, then compares predicted and observed recall within each group. For both, lower is better.

Reading the rows. Error figures are from the benchmark's results table “without same-day reviews”, row “FSRS-6”. The percentages are from its head-to-head chart, where the entry is labelled “FSRS-6 recency” — FSRS-6 trained with more weight on recent reviews. FSRS-7 and the neural networks are more accurate than FSRS-6; Dyscendi currently ships FSRS-6.

Straight talk

What this doesn't prove.

The research is solid. The step from it to “this app will raise your grades” is one we haven't measured, and won't pretend to have.

Four things to keep in mind

We have run no study of our own. Everything above was measured by other people, on other learners, often in a lab. None of it is a measurement of Dyscendi.
The benchmark scores prediction, not learning. It shows FSRS-6 forecasts recall more accurately than several alternatives. It does not show, by itself, that people who use it learn more.
The benchmark's FSRS-6 is fitted to each user. Its parameters are optimised on that person's review history. Dyscendi runs FSRS-6 with its standard parameters and doesn't refit them yet, so its predictions are less tailored than the ones scored here.
A schedule only works if you show up. And a card can only be as good as the notes it came from, which is why Dyscendi makes you approve each one.

References

Every source, in full.

Murre, J. M. J., & Dros, J. (2015). Replication and analysis of Ebbinghaus' forgetting curve. PLOS ONE, 10(7), e0120644. https://doi.org/10.1371/journal.pone.0120644
Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380. https://doi.org/10.1037/0033-2909.132.3.354
Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., & Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095–1102. https://doi.org/10.1111/j.1467-9280.2008.02209.x
Roediger, H. L., III, & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
Karpicke, J. D., & Roediger, H. L., III (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968. https://doi.org/10.1126/science.1152408
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
Ye, J., Su, J., & Cao, Y. (2022). A stochastic shortest path algorithm for optimizing spaced repetition scheduling. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4381–4390). https://doi.org/10.1145/3534678.3539081
Open Spaced Repetition. SRS Benchmark [data and code]. GitHub. Consulted 12 October 2026. https://github.com/open-spaced-repetition/srs-benchmark

Found a mistake, or a study we should read? Write to question@dyscendi.com.

Get the app

Try it on your next chapter.

Free to start, on iPhone and Android.