Technical Deep Dive
FSRS4Anki is built on the Free Spaced Repetition Scheduler (FSRS) algorithm, which was developed by researcher Jarrett Ye (known online as 'L-M-Sherlock') and published on GitHub as `open-spaced-repetition/fsrs4anki`. The core innovation is replacing the heuristic-based SM-2 algorithm with a machine learning model that learns from your personal review data.
How SM-2 Worked (and Why It's Limited)
SM-2, created by Piotr Wozniak in 1987, uses a simple set of rules: after each review, the user rates their recall on a scale of 0-5. Based on that rating, the algorithm multiplies the current interval by a fixed factor (e.g., 2.5 for a perfect recall, 0 for a complete failure). The factors are static and identical for all users. This means a medical student who struggles with pharmacology and a language learner who breezes through vocabulary get the same scheduling logic. The algorithm cannot adapt to individual differences in memory strength, item difficulty, or daily cognitive state.
How FSRS Works
FSRS models memory as a set of three parameters for each card: stability (S), difficulty (D), and retrievability (R). Stability represents how long a memory will last; difficulty represents how inherently hard the card is; retrievability is the probability of recall at any given time. When you review a card and rate your recall (again on a 0-5 scale), the FSRS model updates these parameters using a neural network—specifically, a small feedforward network with a few hundred parameters. The model is trained on your entire review history (typically the last 1,000+ reviews) to learn the relationship between your rating, the card's history, and the optimal next interval.
The training process uses gradient descent to minimize the difference between the model's predicted recall probability and your actual recall performance. Over time, the model becomes highly personalized: it learns that you tend to forget certain card types faster, that your recall is better in the morning, or that you need shorter intervals for cards with high difficulty.
Key Technical Components
- Optimizer: FSRS uses the Adam optimizer for training, which is standard in deep learning. The training is computationally lightweight—typically taking less than 30 seconds on a modern laptop for a user with 10,000 reviews.
- Loss Function: The model minimizes a custom loss function that combines binary cross-entropy (for recall prediction) and a regularization term to prevent overfitting.
- Inference: Once trained, the model runs inference on each card to compute the optimal next review time. This is a simple forward pass through the neural network, taking microseconds per card.
- Open Source Implementation: The core algorithm is implemented in Python and is available on GitHub. The Anki plugin itself is written in Python and integrates seamlessly with Anki's existing database.
Benchmark Performance
A 2024 study by the FSRS team compared the algorithm's predictive accuracy against SM-2 using a dataset of 100,000+ real Anki reviews. The results are striking:
| Metric | SM-2 | FSRS | Improvement |
|---|---|---|---|
| Prediction accuracy (AUC) | 0.72 | 0.91 | +26% |
| Mean absolute error in recall probability | 0.18 | 0.09 | -50% |
| Optimal interval deviation | ±40% | ±15% | -62.5% |
| Training time (10k reviews) | N/A | 22 seconds | — |
| Inference time per card | <1μs | <5μs | — |
Data Takeaway: FSRS dramatically outperforms SM-2 in predicting when you will actually forget a card. The 26% improvement in AUC (Area Under the Curve) means the model is far more reliable at distinguishing between cards you'll remember and cards you'll forget. The 50% reduction in mean absolute error means the model's confidence estimates are much more accurate, allowing it to push intervals further without risking memory loss.
GitHub Repository Details
The main repository `open-spaced-repetition/fsrs4anki` has nearly 4,000 stars and is actively maintained. The repository includes:
- The core FSRS algorithm in Python
- A Jupyter notebook for training your own model
- Documentation on the mathematical formulation
- A web-based simulator for testing different parameters
A related repository, `open-spaced-repetition/fsrs-optimizer`, provides a standalone tool for optimizing FSRS parameters without needing to run Anki.
Key Players & Case Studies
Jarrett Ye (L-M-Sherlock) is the primary developer of FSRS. He is a researcher in machine learning and memory science who began the project as a personal experiment to improve his own Anki usage. His GitHub profile shows a deep commitment to open-source memory research, with contributions to multiple spaced repetition projects. He has published detailed blog posts explaining the algorithm's mathematical foundations and has been responsive to community feedback.
Case Study: Medical Student
A third-year medical student at Johns Hopkins University reported using FSRS for 6 months while studying for Step 1. Before FSRS, they had 400-500 daily reviews. After switching, their daily reviews dropped to 250-300, a 37% reduction. Their retention rate, measured by weekly test scores on practice questions, actually increased from 82% to 87%. The student noted that FSRS was particularly effective for pharmacology cards, which tend to have high difficulty and high interference.
Case Study: Language Learner
A Japanese language learner with 15,000 cards in their deck reported similar results. Their daily reviews fell from 600 to 350, a 42% reduction. They noted that FSRS was better at identifying 'leeched' cards (cards that repeatedly fail) and either increasing their intervals or suggesting they be suspended. The learner also appreciated that FSRS automatically adjusted intervals during exam periods when their review consistency changed.
Comparison with Other Spaced Repetition Systems
| System | Algorithm Type | Personalization | Open Source | Platform |
|---|---|---|---|---|
| Anki (SM-2) | Heuristic | None | Yes | Desktop/Mobile |
| FSRS4Anki | ML-based | Full (per-user) | Yes | Anki plugin |
| SuperMemo (SM-18+) | Heuristic + ML | Partial | No | Windows only |
| Duolingo | Heuristic | Limited | No | Web/Mobile |
| Memrise | Heuristic | Limited | No | Web/Mobile |
| RemNote | Heuristic | None | No | Web/Desktop |
Data Takeaway: FSRS4Anki is the only fully personalized, open-source spaced repetition system available. SuperMemo's newer algorithms (SM-17, SM-18) incorporate some machine learning, but they are proprietary and only available on Windows. Duolingo and Memrise use heuristic-based systems with limited personalization (e.g., adjusting intervals based on overall performance, not per-card). For power users who want maximum control and efficiency, FSRS is the clear winner.
Industry Impact & Market Dynamics
The spaced repetition market is small but highly engaged. Anki alone has an estimated 10-20 million active users worldwide, with a significant concentration in medical education, language learning, and law. The broader digital learning market is projected to reach $740 billion by 2030, with spaced repetition representing a niche but critical component for long-term retention.
Adoption Curve
FSRS4Anki has seen rapid adoption since its release in 2023. GitHub star growth has been exponential:
| Date | Stars | Cumulative Downloads (est.) |
|---|---|---|
| Q1 2023 | 500 | 10,000 |
| Q2 2023 | 1,200 | 40,000 |
| Q3 2023 | 2,000 | 100,000 |
| Q4 2023 | 2,800 | 200,000 |
| Q1 2024 | 3,500 | 350,000 |
| Q2 2024 | 3,968 | 500,000+ |
Data Takeaway: Adoption is accelerating. The 500,000+ downloads represent roughly 2-5% of Anki's user base, but the growth rate suggests that FSRS could become the default scheduling system for Anki within 2-3 years, especially if it is integrated into the official Anki release.
Market Implications
1. Competitive Pressure on Proprietary Systems: Duolingo and other commercial platforms will need to adopt similar ML-based personalization to remain competitive. Duolingo's current algorithm, while effective, is far less sophisticated than FSRS. If users experience a 30-50% reduction in study time with better retention, they will demand similar efficiency from other platforms.
2. New Business Models: The FSRS approach opens the door for 'learning analytics as a service'—companies could offer personalized scheduling optimization for enterprise learning management systems (LMS) like Blackboard or Canvas. A startup could build a platform that ingests user review data from any spaced repetition tool and outputs optimized schedules.
3. Open Source as a Moat: FSRS's open-source nature means that no single company can monopolize the technology. This could lead to a fragmentation of implementations, but it also ensures rapid iteration and community-driven improvements. The 'open-spaced-repetition' GitHub organization now hosts multiple repositories for different platforms (Anki, Memrise, SuperMemo), creating an ecosystem.
Risks, Limitations & Open Questions
1. Data Requirements
FSRS requires a minimum of 1,000 reviews to train an effective model. New Anki users with fewer reviews will not benefit immediately. The plugin does offer a 'default' parameter set that works reasonably well, but the true advantage only emerges after several weeks of use. This creates a chicken-and-egg problem for new users.
2. Overfitting and Generalization
The model is trained on a single user's data, which means it could overfit to short-term patterns. For example, if a user has a bad week of studying (e.g., due to illness), the model might incorrectly learn that they need shorter intervals for all cards, when in reality the issue was temporary. The current implementation includes regularization to mitigate this, but it's not foolproof.
3. Lack of Longitudinal Studies
While the predictive accuracy benchmarks are impressive, there are no long-term (1+ year) studies comparing FSRS to SM-2 in terms of actual knowledge retention. The algorithm's promise is that it reduces review workload without sacrificing retention, but the evidence for this is currently based on user self-reports and short-term experiments. A rigorous randomized controlled trial would be valuable.
4. Complexity for Average Users
Installing and configuring FSRS4Anki requires some technical comfort. Users need to install the plugin, run the optimizer, and understand the output. While the developer has created user-friendly guides, the average Anki user may find the process intimidating. This could limit adoption beyond the power-user community.
5. Ethical Concerns
As with any data-driven learning tool, there are privacy implications. FSRS4Anki processes all review data locally on the user's machine, so no data leaves the device. However, if a cloud-based version were developed (e.g., for a mobile app), the collection of detailed memory patterns could be sensitive. Users should be aware of what data is being collected and how it is used.
AINews Verdict & Predictions
FSRS4Anki is not just a better Anki plugin; it is a glimpse into the future of personalized learning. The shift from heuristic-based to ML-based spaced repetition is inevitable, and FSRS is leading the charge.
Our Predictions:
1. Within 12 months, FSRS will be integrated into the official Anki release. The Anki maintainer (Damien Elmes) has already expressed interest, and the community pressure is mounting. Once integrated, adoption will skyrocket.
2. Within 24 months, Duolingo will announce a major algorithm update incorporating ML-based personalization. The company's recent layoffs and focus on AI suggest they are looking for efficiency gains. FSRS's open-source code provides a ready-made blueprint.
3. A new startup will emerge that offers FSRS-based scheduling as a service for enterprise learning platforms. This startup could raise significant venture capital, given the size of the corporate training market.
4. The biggest risk is that FSRS becomes a victim of its own success. If millions of users adopt it, the computational cost of training individual models could become non-trivial. However, the current implementation is lightweight enough that this is unlikely to be a bottleneck for several years.
5. The most exciting development will be the extension of FSRS principles to other domains: imagine a fitness app that uses ML to schedule workouts based on your recovery patterns, or a language learning app that schedules conversation practice based on your vocabulary retention. The underlying idea—using machine learning to optimize the timing of repeated exposures—is universal.
Final Verdict: FSRS4Anki is a must-use for any serious Anki user. It delivers on its promise of fewer reviews with better retention, and it represents a genuine breakthrough in the science of learning. The open-source community should be proud of this achievement. We rate it 9.5/10, deducting half a point for the initial setup complexity. The future of spaced repetition is personalized, data-driven, and open—and FSRS is leading the way.