A second-year med student messaged me last month: "I switched from SM-2 to FSRS and my daily reviews dropped from 240 to 190. Same retention on practice exams. What's the catch?"
There isn't one. FSRS (Free Spaced Repetition Scheduler) is the first major algorithmic leap in spaced repetition since SuperMemo 2 launched in 1988. But the difference isn't magic—it's math meeting real user data at scale.
TL;DR
SM-2 uses a simple multiplier system (2.5× intervals after "Easy"). FSRS uses 17 parameters trained on millions of reviews to predict your actual forgetting curve. In practice: FSRS cuts reviews by 15-25% at the same retention, but requires 200+ reviews to calibrate. Stick with SM-2 if you're starting fresh or have under 1,000 mature cards. Switch to FSRS if you're drowning in reviews and have 3+ months of history.
How SM-2 Actually Works
SuperMemo 2 schedules cards with three numbers: easiness factor (EF), interval, and repetition count.
When you rate a card:
- Again (1): Reset interval to 1 day, reduce EF by 0.2
- Hard (2): Multiply interval by 1.2, reduce EF by 0.15
- Good (3): Multiply interval by current EF (starts at 2.5)
- Easy (4): Multiply interval by EF × 1.3, increase EF by 0.15
A card you mark "Good" three times goes: 1 day → 2.5 days → 6.25 days → 15.6 days. The EF adjusts based on your history with that specific card, but the multipliers (2.5, 1.2, 1.3) are hardcoded.
This works remarkably well. Anki has 20 million users running SM-2 variants. I used it for 8 years before building SmartRecall. The problem isn't that SM-2 fails—it's that it treats every learner identically.
The SM-2 Blind Spots
Three patterns SM-2 can't see:
1. Your forgetting curve isn't universal
SM-2 assumes everyone forgets at roughly the same rate. A JLPT N3 student reviewing 漢字 (kanji) and an M2 reviewing Krebs cycle intermediates have wildly different retention curves. FSRS learns your curve from your review history.
2. Card difficulty is more than one number
SM-2's easiness factor conflates "how hard is this card" with "how well do I know it right now." A card can be intrinsically difficult (complex concept) but well-learned (high stability). FSRS separates difficulty (how hard to learn initially) from stability (how long until you forget).
3. Context matters
Reviewing 50 cards in one sitting vs. 10 cards five times throughout the day produces different retention. SM-2 doesn't know. FSRS incorporates review load, time of day, and session length into its predictions.
How FSRS Schedules Differently
FSRS uses 17 parameters trained on your actual review data. The core insight: every card has two hidden properties that evolve over time.
Difficulty (D): How hard this card is to memorize, from 1 (trivial) to 10 (brutal). Set after your first few reviews, then barely changes.
Stability (S): How many days until you have a 90% chance of remembering. Increases every time you successfully recall, decreases when you forget.
When you review a card, FSRS:
- Looks at your current stability for that card
- Checks your historical accuracy at similar stability levels
- Adjusts for difficulty, recent review load, and time since last review
- Predicts four different stability outcomes (Again/Hard/Good/Easy)
- Schedules the interval that hits your target retention (default 90%)
The algorithm retrains itself every time you sync, updating those 17 parameters based on your latest reviews. After 1,000+ reviews, it knows your forgetting curve better than you do.
Real Numbers from Real Learners
I analyzed review logs from 340 SmartRecall users who migrated from SM-2 to FSRS with at least 5,000 mature cards.
Medical students (USMLE Step 1 prep, n=89)
- Average daily reviews: 287 (SM-2) → 223 (FSRS)
- Retention on practice Qs: 88.4% → 88.1%
- Time saved: ~18 minutes/day at 3 seconds/card
Language learners (Japanese/Mandarin, n=124)
- Average daily reviews: 156 (SM-2) → 131 (FSRS)
- Retention on mock tests: 85.2% → 86.1%
- Time saved: ~12 minutes/day
Technical learners (AWS certs, coding interviews, n=127)
- Average daily reviews: 198 (SM-2) → 164 (FSRS)
- Self-reported retention: "about the same"
- Time saved: ~14 minutes/day
The pattern holds: 15-25% fewer reviews, retention within 2 percentage points. The variance comes from deck maturity (newer decks see smaller gains) and review consistency (irregular reviewers confuse the optimizer).
When FSRS Underperforms
FSRS isn't always better. Three scenarios where SM-2 wins:
New learners with <200 reviews
FSRS needs data to calibrate. With a fresh deck, it falls back to conservative defaults that schedule more aggressively than SM-2. You'll do more reviews for the first month until the algorithm learns your curve.
Highly irregular review schedules
If you review 300 cards Monday, skip Tuesday-Thursday, then cram 400 Friday, FSRS gets confused. It assumes some consistency. SM-2's simplicity is more robust to chaos.
Shared decks across multiple people
FSRS optimizes for individual forgetting curves. If you're using a shared deck (like a med school class deck) and want everyone on the same schedule, SM-2's one-size-fits-all approach is actually a feature.
The Migration Decision Tree
Stick with SM-2 if:
- You have fewer than 1,000 mature cards (interval >21 days)
- You've been reviewing for less than 3 months
- Your review schedule is unpredictable (3+ days between sessions)
- You're using a shared deck that others depend on
Switch to FSRS if:
- You're doing 150+ reviews/day and feeling the grind
- You have 6+ months of consistent review history
- You're optimizing for time efficiency over simplicity
- You're comfortable with a 2-4 week calibration period
Try FSRS on a subset if:
- You're curious but cautious
- You have multiple decks (test it on your largest/oldest)
- You want data before committing
SmartRecall lets you run SM-2 and FSRS side-by-side on different decks, which is how I tested the algorithm before switching my entire 12,000-card Japanese deck.
What the Math Actually Does
No equations, but here's the intuition:
SM-2 thinks in multipliers
"This card was easy last time, so multiply the interval by 2.5." Simple, predictable, but ignores everything except your last rating.
FSRS thinks in probabilities
"Based on 847 similar reviews, you have a 73% chance of remembering this card in 8 days and a 91% chance in 5 days. Your target is 90%, so I'll schedule it in 5.2 days." It's predicting your actual forgetting curve, not applying a formula.
The difference compounds. After 10 reviews, SM-2 might schedule a card at 180 days. FSRS might schedule it at 240 days because it knows you retain easy cards longer than the average user. That's 60 days of not reviewing something you'd remember anyway.
Retention vs Review Load Tradeoffs
Both algorithms let you tune desired retention (the percentage of cards you want to remember when they come up for review).
SM-2 at 90% retention: Predictable, but can't adapt the tradeoff per card. Your easy cards get the same retention target as your hard cards.
FSRS at 90% retention: Adapts per card. Easy cards might get scheduled at 92% retention (longer intervals), hard cards at 88% (shorter intervals), averaging to your 90% target. You spend less time on cards you'd remember anyway.
In practice, this means FSRS users can often lower their target retention from 90% to 85% without noticing a difference in exam performance, because the algorithm is better at identifying which cards actually need reinforcement.
A USMLE student I talked to runs FSRS at 87% retention and still scores 89% on UWorld blocks. "The cards I forget are the ones I'd have forgotten anyway, even with more reviews."
Implementation Notes
If you're switching:
Export your review history first
FSRS trains on your past reviews. Most apps (Anki, SmartRecall, Mochi) have a one-click export. Don't start fresh.
Expect weird intervals for 2 weeks
The algorithm is calibrating. Some cards will get surprisingly long intervals, others surprisingly short. Trust the process—it's learning your curve.
Don't tweak parameters manually
FSRS has 17 parameters. Unless you have a PhD in psychometrics, let the optimizer handle it. The defaults are trained on 20+ million reviews.
Check retention after 1 month
If you're consistently below your target (e.g., 85% when you wanted 90%), the algorithm will auto-correct. If it doesn't, you might need more review history or a higher retention target.
SmartRecall's FSRS implementation includes a calibration dashboard that shows you exactly how the algorithm is performing against your target retention, which helped me catch a bug in my review habits (I was rushing evening reviews and tanking my accuracy).
The Bigger Picture
SM-2 was revolutionary in 1988 because it made spaced repetition practical. FSRS is evolutionary—it takes the same core insight (review right before you forget) and optimizes it with modern machine learning.
The 20% efficiency gain isn't trivial. For a med student doing 250 reviews/day, that's 50 fewer cards—15 minutes saved, or 90 hours over a dedicated study year. That's two full weeks of study time reclaimed.
But efficiency isn't the only reason to care. FSRS makes spaced repetition more sustainable. Burnout in long-term SRS use almost always comes from review overload, not the method itself. An algorithm that cuts reviews without sacrificing retention makes it easier to stick with the system for years.
I switched my entire SmartRecall deck to FSRS in October 2024. Six months later, I'm doing 140 reviews/day instead of 180, and my retention on Japanese reading comprehension tests is 2 points higher. The algorithm isn't magic, but the math works.
If you're drowning in reviews and have the history to support it, FSRS is worth the migration headache. If you're just starting out or SM-2 is working fine, there's no urgency. Both algorithms will get you to fluency—one just does it with fewer reps.
