Mibuki's own spaced-repetition (SRS) model. On three public benchmarks it beats the default settings of the industry-standard FSRS-7 by a wide margin — delivering strong personalized accuracy before any per-user optimization.
← Back to LangDictPublished 21 June 2026 · Last updated 28 July 2026 · Benchmark run June 2026 · Model version Lexa 3.0
Your memory habits are shared across cards — so your other cards are the best hint.
The heart of a spaced-repetition scheduler is predicting how well you'll recall a card, which decides when to review it. The industry-standard FSRS personalizes by fitting parameters per user (per-user fitting) — powerful, but it needs tuning for each user.
Lexa takes the opposite route. Traits like how fast you forget, how you rate, and your review rhythm are shared across your cards. Lexa reads not just the target card but your other cards' review history as in-context input, so a single shared model personalizes with no per-user tuning.
In other words — instead of learning “other users' tendencies,” it reads how you forget directly from the rest of your own cards. That is what's new about Lexa.
It reads your other cards' history, estimates how well each card is retained, and sets the next review date.
Lexa beats the out-of-the-box (default) FSRS-7 on all three public benchmarks — before any per-user optimization.
Memory-prediction score (1 − Log Loss, higher is better)
FSRS moves quickly. What is published here is the comparison as of the dates above; a newer FSRS may score differently. The baseline is the default configuration of FSRS-7, and the baseline figures below were re-computed from the stored run in the LangDict repository (research/srs-benchmark/result) and match.
| Dataset | FSRS-7 (default) | Lexa 3.0 |
|---|---|---|
| Anki | 0.646 (.3543) | 0.682 (.3184) |
| Duolingo | 0.577 (.4231) | 0.672 (.3280) |
| maimemo | 0.488 (.5121) | 0.538 (.4621) |
Lexa needs no per-user tuning or training and is accurate without a per-user fitting phase. Out of the box — before either model is fitted to you — it already beats FSRS-7’s default settings on all three datasets. The margin is not uniform: +0.095 on Duolingo, +0.050 on maimemo, +0.036 on Anki.
Evaluated on three real-world datasets.
~727M reviews / 10,000 users
Standard flashcards
~226M reviews / vocabulary
Chinese vocabulary app
~13M reviews / session-based
HLR dataset
Metric: per-user weighted Log Loss (lower is better).
Accuracy isn't just a benchmark number — it shapes your learning directly.
Lexa pinpoints the moment you're about to forget. Too early and reviews are boring; too late and you have to relearn. Lexa narrows that window. In the app, the report shows how much it cut the reviews scheduled over the last 7 days. That number comes from replaying the same study history through SM-2 and comparing — it is an estimate of scheduling load, not a measurement of learning outcomes.
No parameters to tune and no waiting for per-user learning. Just open the app and study — it schedules to you from the start.
Personalization runs entirely on your device. Privacy and accuracy together.
We keep experimenting on our own benchmark pipeline and continuously improve the model we ship.
open-spaced-repetition. FSRS: Free Spaced Repetition Scheduler. GitHub.
open-spaced-repetition. srs-benchmark: Benchmark of Spaced Repetition Algorithms. GitHub.
open-spaced-repetition. anki-revlogs-10k: Anki review logs from 10,000 users. Hugging Face.
Settles, B., & Meeder, B. (2016). A Trainable Spaced Repetition Model for Language Learning. Proceedings of the 54th Annual Meeting of the ACL.
Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Leipzig: Duncker & Humblot.