Anki has shipped a second scheduler since version 23.10, and it is switched off until you turn it on. That scheduler is FSRS. It is easy to keep reviewing on the older SM-2 settings without knowing the switch exists, or to turn FSRS on and then stare at a "desired retention" field with no idea what number to type.
Both problems come from the same gap: FSRS is easy to enable and hard to picture. Once you can see the three numbers it keeps for every card, the settings stop looking arbitrary.
The short version: FSRS (Free Spaced Repetition Scheduler) is an open-source algorithm that estimates, for each card, how likely you are to recall it today, and schedules the next review for the day that probability falls to a target you choose. Anki includes it from version 23.10, and its optimizer can fit the model to your own review history. Memly also schedules with FSRS, but it uses one shared set of parameters for every user and does not run that optimization.
What is FSRS?
Spaced repetition schedules reviews at widening intervals, so you revisit material shortly before you would forget it. The hard part is deciding when that is for one particular card, and FSRS answers it with a model of memory instead of a fixed ladder of intervals.
According to the ABC of FSRS page in the project's wiki, FSRS was developed by Jarrett Ye, a research engineer at MaiMemo Inc. It descends from MaiMemo's DHP model, a variant of the three-component DSR model (difficulty, stability, retrievability). The current release in Anki, FSRS-6, uses 21 parameters in its formulas for those three quantities (The Algorithm, checked October 11, 2026).
The parameters are where personalization lives. Until a user has enough reviews, the wiki says, the default parameters are used; they were found by running the optimizer on several hundred million reviews from about 10,000 users. After that, an app can fit the 21 parameters to the user's own history. Whether it does is the main difference between FSRS in Anki and in Memly.
Difficulty, stability, retrievability: the three numbers FSRS tracks
Every card in FSRS carries a memory state with three parts, defined in the ABC of FSRS:
- Difficulty (D): how hard it is to raise the card's stability after a review. It is bounded between 1 and 10. A high-difficulty card gains less stability from each successful review, so its intervals grow more slowly.
- Stability (S): the time, in days, for the probability of recall to fall from 100% to 90%.
- Retrievability (R): the probability that you can recall the card at a given moment. It depends on the time since the last review and on the stability.
Stability is measured in days, not confidence. A card at stability 40 has a 90% chance of being recalled forty days after its last review. The wiki's own example is a card at S = 365: a full year passes before its probability of recall drops to 90%.
The FSRS-6 forgetting curve
Retrievability is calculated from stability with a power function. In FSRS-6 the formula is:
R(t, S) = (1 + FACTOR × t / S) ^ (−w20), where FACTOR = 0.9 ^ (−1 / w20) − 1, so that R = 90% when t = S.
Here t is the number of days since the last review and w20 is the decay parameter. Earlier versions fixed the decay as a constant. FSRS-6 is the first version that makes it trainable, so that, in the words of the SRS benchmark README, "the shape of the curve is different for different users." The default value of w20 is 0.1542, and the reference TypeScript implementation, ts-fsrs, keeps it between 0.1 and 0.8 when parameters are clipped. For most users, it is optimized to below 0.2. A curve unique to you still requires fitting the parameters to your own reviews.
From the curve to the next review date
The scheduler solves the same equation backwards. It sets R to your desired retention and finds the t at which the curve reaches it. At the default desired retention of 90%, the next interval equals the card's stability, which is why a stability-40 card comes back in about forty days. Raise the target and the interval shrinks; lower it and the interval stretches.
After each review, FSRS updates difficulty and stability from your grade (Again, Hard, Good or Easy) and from how far the card had decayed. A successful recall at low retrievability raises stability more than one at high retrievability: the spacing effect, written into the model. Same-day reviews have their own formula, controlled by w17, w18 and w19, and the wiki notes that in practice it should not lower stability on any passing grade.

How FSRS evolved: from Ebbinghaus to FSRS-6
The idea that memory decays along a curve is about 140 years old, dating to Ebbinghaus's 1885 study. What changed recently is that the curve is now fitted to large sets of real review logs. Dates below come from the FSRS release tags and the Anki release notes.
| Year | Step | What changed |
|---|---|---|
| 1885 | Ebbinghaus's forgetting curve | First experimental measurements of forgetting over time |
| 1987 | SM-2 (SuperMemo) | First used in SuperMemo 1.0 for DOS; the first computer algorithm for computing a spaced repetition schedule |
| 2022 | FSRS v3 | Custom scheduling script for Anki; exponential forgetting curve |
| 2023 | FSRS v4 and 4.5 | Switch to a power-function forgetting curve; v4 was integrated into Anki 23.10 |
| 2024 | FSRS-5 | Uses same-day reviews to refine the next prediction |
| 2025 | FSRS-6 | 21 parameters, trainable forgetting-curve shape; shipped in Anki 25.07 |
The SRS benchmark already lists FSRS-7, designed for fractional interval lengths, with eight optimizable parameters in its forgetting curve. We found no Anki release note announcing it (checked October 11, 2026), so FSRS-6 remains the version in Anki, and the one Memly runs.
The FSRS development team is also considering more specialized versions, such as FSRS-F (which factors in fatigue) and FSRS-S (which specializes in short-term memory). Neither is finished.
How to turn on FSRS in Anki
Turning FSRS on takes a minute; the Anki manual's FSRS section (checked October 11, 2026) gives a short guide. First make sure every Anki client you use supports it: Anki 23.10, AnkiMobile 23.10, AnkiWeb, and AnkiDroid 2.17 or later. If one does not, the manual warns that things will not work correctly.
- Open the deck options and find the "FSRS" section at the bottom of the page. Enable it. FSRS can only be enabled globally; you cannot turn it on for some presets and off for others.
- Keep your learning and relearning steps shorter than one day, so they can be completed on the same day. The manual suggests steps such as 10m or 30m.
- Click "Optimize" under "FSRS parameters". If Anki says the parameters already appear to be optimal, that is fine.
- Choose a desired retention. The default is 90%, which the manual describes as a good balance of retention and workload.
- Leave "Reschedule cards on change" off when you first switch from SM-2. With it off, future reviews use FSRS but your current due dates stay put; with it on, a large number of cards can become due at once.
One habit breaks FSRS: pressing Hard when you actually forgot. FSRS treats Hard as a pass, so the manual warns that intervals become unreasonably long. Use Again whenever you failed to recall.
How many reviews do you need before optimizing?
Older guides, including an earlier version of this article, said to wait for about 1,000 reviews before optimizing, with partial optimization possible from around 400. The Anki FSRS FAQ (checked October 11, 2026) now says the 1,000-review requirement applied only to earlier versions: from Anki 24.06.3 the optimizer can be used with any number of reviews.
Few reviews still mean a weaker fit. The manual's health check lists "less than a few hundred" reviews as a common reason FSRS may not adapt well. For timing, the FAQ says once a month is more than enough, or, as a finer rule, optimize each time your review count doubles: after 100, 200, then 400 reviews.
One preset or several?
Parameters and desired retention are set per preset, so decks with dissimilar material can learn different parameters. The FAQ's example is Japanese and geography: two presets. Two Japanese vocabulary decks can share one. The manual also notes that in the latest version of Anki you can set different desired retention values for different decks within the same preset. That is how a bar-exam candidate could hold constitutional law at 95% and commercial law at 85% while both use the same fitted parameters.
FSRS vs SM-2: what actually changes
Anki calls SM-2 its legacy algorithm. It stores an ease factor per card and multiplies the last interval by it. Nothing in SM-2 estimates the probability that you will remember a card, so there is no number to aim at and no way to fit the schedule to your history.
| Aspect | SM-2 (Anki's legacy scheduler) | FSRS |
|---|---|---|
| What it stores per card | Interval and ease factor | Difficulty, stability, and retrievability |
| What you set | Starting ease, graduating interval, easy bonus | Desired retention (default 90%) |
| Fitting to your history | None | Optimizer fits 21 parameters (in Anki) |
| First intervals | Usually shorter | Often 3–5 days for Good on a new card, per the Anki FAQ |
| Mature cards | Intervals keep multiplying | More conservative than SM-2, per the Anki FAQ |
What does the FSRS project claim about efficiency? The ABC of FSRS says that with FSRS users need 20–30% fewer reviews than with SM-2 to reach the same retention level, and it labels that figure as based on simulation results. It also says FSRS is better than SM-2 even with default parameters, and that it handles cards reviewed late, after a break of weeks or months, much better. The Anki manual phrases the benefit more modestly: by estimating more accurately how much you are likely to forget, FSRS "can help you remember more material in the same amount of time." Both are statements by the people who build and ship FSRS, and the 20–30% is a simulation, not a measured result for any one learner.
What fewer reviews mean in study hours
Fewer reviews add up over a year of daily study. An earlier version of this article illustrated the difference with 10,000 English vocabulary words:
| Metric | Traditional method | FSRS 6.0 |
|---|---|---|
| Time required | 2 hours/day × 18 months | 1.5 hours/day × 12 months |
| Total study time | Approx. 1,080 hours | Approx. 540 hours |
| Time reduction | – | Approx. 50% |
The line to read is the total: the same 10,000 words in half the hours. This is an illustrative estimate, not data from Memly users, and it shortens both the daily time and the number of months, which is why it goes beyond the 20–30% review figure. The ts-fsrs simulation further down states its assumptions in full.
What the benchmark measures
The project also runs a public SRS benchmarkthat scores how well each algorithm predicts whether a review will succeed. Its earlier dataset came from about 20,000 Anki users and 1.7 billion reviews; the current one uses about 10,000 users and 727 million reviews. Log loss and RMSE (bins) measure prediction error, so lower is better. These are the rows for the FSRS versions in the table that excludes same-day reviews (9,999 collections, checked October 11, 2026):
| Algorithm | Parameters | Log loss | RMSE (bins) |
|---|---|---|---|
| FSRS-7 | 34 | 0.3401 | 0.0634 |
| FSRS-6 | 21 | 0.3460 | 0.0653 |
| FSRS-5 | 19 | 0.3561 | 0.0742 |
| FSRS-4.5 | 17 | 0.3625 | 0.0764 |
| FSRS v3 | 13 | 0.4364 | 0.1097 |
The benchmark also reports how often one algorithm beats another per collection. In the current superiority chart, FSRS-6 trained with recency weighting had lower log loss than FSRS-5 in 84.2% of collections and than FSRS-4.5 in 92.4%.
An earlier version of the benchmark, run on data from about 20,000 Anki users, also included SM-2. These are its results as reported in the earlier version of this article; the rightmost column shows the share of users for whom FSRS-6 predicted recall more accurately. The current README no longer has an SM-2 row, so these figures cannot be checked against it.
| Algorithm | Log loss | RMSE | Users where FSRS-6 wins |
|---|---|---|---|
| FSRS-6 | 0.30 | 0.052 | – |
| FSRS-5 | 0.33 | 0.058 | 88.2% |
| FSRS-4.5 | 0.34 | 0.061 | 92.1% |
| SM-2 | 0.41 | 0.089 | 99.6% |
Read the last column. An average gain can hide a minority who did worse; at 99.6% against SM-2, you are unlikely to be one.
How desired retention changes your review load
The Anki manual calls desired retention the most important setting in FSRS. Higher retention means shorter intervals and more reviews per day. The manual warns that the workload climbs steeply above 90%, that above 97% it can be overwhelming, and recommends staying below 97%.
How steep is that climb? The interval formula answers it for a single card. For a card with stability 40 days and the default FSRS-6 decay, the next interval at each target is:
| Desired retention | 80% | 85% | 90% | 93% | 95% | 97% |
|---|---|---|---|---|---|---|
| Next interval (S = 40 days) | 133 days | 76 days | 40 days | 25 days | 16 days | 9 days |
Moving from 90% to 95% cuts that interval by more than half. A whole deck behaves differently from one card, because reviews at higher retrievability also build stability more slowly, so we simulated a deck with ts-fsrs.
For a current settings walkthrough and a separate reproducible example, see Anki FSRS settings at 80%, 90% or 95%. It includes real controls and a worksheet for checking your own workload.
A desired-retention target and an observed pass rate are different quantities. Our flashcard retention-rate examples explain the denominator and why a short session can differ from a probability target without proving the model changed.
A 10,000-card simulation with ts-fsrs
The scenario: a learner adds 10,000 vocabulary cards, 20 a day for 500 days. The assumptions matter more than the output:
- Scheduler: ts-fsrs 5.4.2 with the FSRS-6 default parameters (the same values Memly uses), no interval fuzz, maximum interval 36,500 days.
- Day-level model: same-day learning steps are switched off, so only reviews on separate days are counted.
- The simulated learner's memory follows the FSRS-6 default model exactly. Each review succeeds with the probability the model predicts. This is the model checking itself, not data from real learners.
- Grades are Good on success and Again on failure; every due card is reviewed on its due day. Results are the average of three random seeds.
| Desired retention | Reviews after first study (500 days) | Reviews per day (days 401–500) | Share of reviews failed | Total vs 90% |
|---|---|---|---|---|
| 80% | 36,420 | 93 | 19.9% | 0.63× |
| 85% | 45,492 | 122 | 15.1% | 0.79× |
| 90% | 57,689 | 157 | 10.9% | 1.00× |
| 93% | 71,002 | 193 | 8.5% | 1.23× |
| 95% | 85,467 | 229 | 6.8% | 1.48× |
| 97% | 116,733 | 316 | 4.7% | 2.02× |
The column to read is the last one. Going from 90% to 95% added about half again as many reviews in this model, and 97% roughly doubled them, while the share of failed reviews dropped by a few points each step. Your own numbers will differ: real memory does not follow the default model exactly, and fitted parameters shift the whole table. In Anki, the "Help Me Decide" simulator in the deck options runs this kind of estimate on your actual cards and parameters.

For how this connects to the classic forgetting-curve research, see When to Review What You Learned. If your queue has already grown beyond what you can clear, read Flashcard Review Backlog: How to Restart Without Burning Out.
How Memly uses FSRS
Memly, an AI flashcard app for Web, iOS and Android, schedules every graded review with FSRS-6 through ts-fsrs (its rapid review mode leaves the schedule untouched). Its 21 weights are the published FSRS-6 default parameters, value for value, and the same set applies to every user. Memly does not run the FSRS optimizer on your history.
Anki can fit FSRS to your own review history; Memly cannot. What adapts in Memly is each card's memory state: difficulty and stability are updated from that card's own answers, so a card you keep missing comes back sooner than one you answer easily. The shape of the forgetting curve, w20 included, stays at the default for everyone. Because nothing is fitted, Memly needs no warm-up period of reviews before scheduling works as intended.
Desired retention is set per deck, under "Learning intensity", with four options (checked in the Memly app code on October 11, 2026):
| Option | Desired retention | Description in the app |
|---|---|---|
| Light | 80% | Minimal reviews, for hobby decks |
| Standard (default) | 90% | Memly's default, marked as recommended |
| Solid | 93% | For serious exams and languages |
| Intense | 95% | For must-not-fail material |
A change takes effect from each card's next review rather than rescheduling everything at once, similar to Anki's default. There is no free-form value, so the 85% in the bar-exam example above is not available; the nearest options are 80% and 90%. If you set an exam date on a deck, Memly raises that deck's desired retention over the last 14 days before the exam, reaching the target (95% by default) for the final three days.
| FSRS in practice | Anki | Memly |
|---|---|---|
| Default scheduler | SM-2; FSRS must be enabled | FSRS, no setting needed |
| FSRS version | FSRS-6 from Anki 25.07 | FSRS-6 (ts-fsrs) |
| Parameters | Optimized on your history, per preset | FSRS-6 defaults, shared by all users |
| Desired retention | A value you choose per preset or deck (manual advises below 97%) | 80%, 90%, 93% or 95% per deck |
| Moving cards between them | Memly imports and exports CSV only, so review history does not carry over and FSRS state starts fresh | |
Anki gives a careful user more control and a model fitted to their memory. Memly removes the setup and the decisions, at the cost of that fit. For a broader comparison of the two apps, see Memly vs Anki.

Frequently asked questions about FSRS
Is FSRS better than SM-2?
The FSRS project reports 20–30% fewer reviews for the same retention, from simulation. That is not a guarantee for every learner, and grading Hard on forgotten cards undermines it.
Does Anki use FSRS by default?
No. Anki has offered two algorithms since 23.10, and FSRS has to be enabled in the deck options.
What desired retention should I use?
Start at the default 90%. Raise it for material you cannot afford to miss, and check the extra workload first; in the simulation above, 95% meant about 1.5 times the reviews of 90%.
Can FSRS be used for last-minute cramming?
FSRS is built for long-term retention. The Anki manual notes that repeating a card many times in one day does not significantly add to long-term memory. Close to an exam, a higher desired retention for that deck, or Memly's exam date setting, brings reviews forward without abandoning the schedule. For a sense of how many repetitions a card needs, see How Many Times Do You Need to Review to Memorize Something?.
Choosing your next step
If you already use Anki, the useful step today is small: open the deck options, enable FSRS, click Optimize, and leave desired retention at 90% for a month before touching it. That alone puts the three numbers in this article to work on your own history.
If you would rather not manage presets and parameters, Memly runs FSRS-6 with shared defaults and four intensity options per deck. For the wider picture of how AI and spaced repetition fit together, see the pillar guide What Is AI-Powered Memorization? How It Works, Its Effectiveness and Best Tools. Memly is free to try, no credit card required.
Sources
- Anki Manual: Deck Options, FSRS (checked October 11, 2026)
- Anki: Frequently asked questions about FSRS (checked October 11, 2026)
- Anki: What spaced repetition algorithm does Anki use? (checked October 11, 2026)
- Anki 23.10 release notes and Anki 25.07 release notes (checked October 11, 2026)
- ABC of FSRS and The Algorithm, open-spaced-repetition wiki (checked October 11, 2026)
- open-spaced-repetition/srs-benchmark (checked October 11, 2026)
- fsrs4anki release tags for FSRS version dates (checked October 11, 2026)
- open-spaced-repetition/ts-fsrs (version 5.4.2 used for the simulation)
- SuperMemo Guru: Algorithm SM-2 (checked October 11, 2026)
