FSRS vs SM-2: why the newer algorithm schedules fewer reviews
FSRS vs SM-2 comes down to a fixed formula versus a fitted model. SM-2 is the ease-factor formula Piotr Wozniak wrote for SuperMemo in the 1980s, and it powered Anki for years: simple, deterministic, the same rules for everyone. FSRS (the open-source Free Spaced Repetition Scheduler) instead fits a model of your memory to your actual review history and predicts how likely you are to recall each card on a given day. On the largest public benchmark it schedules more accurately than SM-2 for almost every user. The practical result is fewer reviews for the same memory.
What SM-2 does
SM-2 gives every card an "ease factor" and nudges it up or down each time you grade a review: get a card right and its interval grows by that factor, get it wrong and it resets. It's a clean rule of thumb from before there was much data to fit against, and it works. But it treats every card and every person with the same arithmetic, and it never asks how likely you actually are to remember. It just multiplies.
What FSRS does
FSRS models memory with three quantities: how hard a card is for you, how stable the memory currently is, and how retrievable it is right now. It fits those to your own review log, then schedules each card to hit a retention target you choose (say, 90%). Because it predicts a probability of recall rather than following a fixed multiplier, it can push easy cards much further out and hold difficult ones closer, tuned to how you specifically forget.
What the benchmark shows
The open-spaced-repetition benchmark is the honest place to settle this: roughly 10,000 real Anki collections and hundreds of millions of reviews, scoring each algorithm on how well its predictions match what learners actually recalled. FSRS predicts more accurately than SM-2 for about 99.6% of users (measured by log loss), and its calibration error (RMSE) is lower than SM-2's across every FSRS version tested. One fair caveat the benchmark authors make themselves: SM-2 was never built to output recall probabilities, so comparing the two requires converting its output, and no comparison is perfectly clean. Even so, every reasonable way of running it lands the same way.
| How they schedule | SM-2 (1980s) | FSRS (modern) |
|---|---|---|
| Core idea | Fixed ease-factor formula | Memory model fitted to data |
| Personalises to your history | No, same rules for all | Yes, fits your review log |
| Predicts your odds of recall | No | Yes, as a probability |
| You pick a retention target | No | Yes (e.g. 90%) |
| Reviews for the same memory | More | Fewer |
What it means for a learner
You don't need to care about the math. You want the same retention for less time at the desk, and that's what the newer scheduler buys you. Two honest notes. First, Anki now ships FSRS too: if you already live in Anki, you can switch it on in the deck settings and get the benefit without changing apps. Second, SM-2 is not broken. Millions of words got learned on it, and for a small, casual deck the difference is modest. FSRS matters most when your reviews pile up and every saved repetition counts.
How we use it
TangoLango runs FSRS by default, with no optimiser to run or settings to tune. You get the modern scheduling as the plumbing, and we build the cards from your own life so there's nothing to type out. If you're comparing methods more broadly, see how spaced repetition works for a language and the wider language-learning methods we build on. Or the short version: does spaced repetition work for language learning?
I don't want learners thinking about ease factors. FSRS quietly means fewer cards in front of you each morning for the same memory. That's the whole pitch: less time, same result.