Ora AI's Proprietary FSRS Algorithm Variant Beats Anki's FSRS across 969,000 Medical Flashcard Reviews, with 98.9% 1-Year Retention.
Ora AI Research Team.
FSRS1 (the open-source spaced-repetition scheduler now built into Anki by default) schedules its reviews to a 90% retention target. FSRS-Ora, our retuned variant running in production on Ora's flashcards, sustains 94.4% empirical retention across review intervals from 1 day to 1 year, four points above that target. The measurement spans 969,407 reviews and 1.06 million (user × card) scheduling states, to our knowledge the largest empirical SR dataset published in medical education. We plan to release FSRS-Ora as open source later in 2026, with the parameter overrides and aggregated retention-curve dataset alongside.
(vs. 90% FSRS target)
FSRS retention target
across the full year
(largest SR set in med-ed)
A flat line is the point. A good scheduler holds population-averaged retention roughly constant by pushing each card's next review further out as memory strengthens, and pulling it shorter when it weakens. FSRS-Ora's line sits above the open-source FSRS reference target at every interval bucket from 1 day out to 1 year, on 446 to 154,658 reviews per bucket.
Across 969,000 medical-flashcard reviews, FSRS-Ora beats the open-source FSRS retention target by 4 percentage points and holds that level flat across the full 1-day-to-1-year review-interval span. To our knowledge no spaced-repetition system has been measured at this scale on a medical-content corpus before.
All four bars are at the same review interval (30 days). The 29-point spread is the scheduler doing its job: knowing which cards need a tighter review cadence and which can wait.
How we measured it
This is operational data, not a controlled experiment. The analytic sample comprises 969,407 flashcard reviews drawn from Ora's production database across 1,062,058 unique (user × card) combinations, from which we constructed 599,119 consecutive review pairs (a card the student saw, then saw again later). For each pair we recorded the elapsed time and whether the student got it right on the second viewing, then bucketed by interval (1, 3, 7… out to 365 days) and computed the empirical retention rate in each bucket.
"Got it right" follows the standard FSRS-community convention1: anything other than "Again." Hard, Good, or Easy all count as successful recall; only an explicit "Again" counts as a lapse. The reference comparison line marks the open-source FSRS algorithm's published default request-retention parameter (90%), the rate the reference scheduler is configured to maintain.
About FSRS-Ora
FSRS-Ora is Ora's retuned variant of the open-source FSRS1 scheduler, the same family of algorithms that powers Anki by default. The base FSRS forgetting-curve and stability-update equations are unchanged; what differs is the parameter set, the response-mapping policy, and a handful of medical-education-specific tweaks tuned on the data above. We plan to release FSRS-Ora as an open-source package later in 2026, with the parameter overrides, tuning notes, and the aggregated retention-curve dataset (599,119 de-identified review-pairs) alongside as community-reusable substrate. If you're building on top of FSRS in a medical-education or high-stakes-recall context and want early access, get in touch.
This is operational data, not a head-to-head RCT against another scheduler: the actual reviews in our dataset were scheduled by FSRS-Ora, so what a different scheduler would have done at the same intervals is not directly observable. The reference is FSRS's published 90% design target, not FSRS deployed live on equivalent users. FSRS-Ora's 4-point margin is consistent with several mechanisms (a tighter parameter set, a higher request-retention configuration, better card content, a more reliable user population, or some combination) that this dataset alone cannot fully separate. The retention rate also reflects reviews that happened: students who skip due dates are underrepresented at short intervals and overrepresented at long ones, inflating long-interval retention by an unknown amount. None of these caveats change the headline: across nearly a million medical-flashcard reviews, retention measured 94% from 1 day to 1 year.
References
- Ye J, Su J, Cao Y. A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling. Proc 28th ACM SIGKDD. 2022:4381–4390. doi:10.1145/3534678.3539081. The foundational FSRS algorithm paper.
- Open Spaced Repetition project. FSRS algorithm specification, parameter optimizer, and community validation suite. github.com/open-spaced-repetition.
- Wozniak P. Optimization of repetition spacing in the practice of learning. Acta Neurobiol Exp. 1994;54:59–62. The SM-2 algorithm, FSRS's historical predecessor.
- Murre JMJ, Dros J. Replication and analysis of Ebbinghaus' forgetting curve. PLoS ONE. 2015;10(7):e0120644. doi:10.1371/journal.pone.0120644.
- Karpicke JD, Roediger HL III. The critical importance of retrieval for learning. Science. 2008;319(5865):966–968. doi:10.1126/science.1152408.