FSRS-6 vs SuperMemo SM-20: Best SRS Algorithm 2026

For two decades, the spaced-repetition engine under most flashcard apps was a lightly-modified version of an algorithm from the late 1980s. In 2026 that finally changed for everyone. The open-source FSRS scheduler reached version 6 and became the default in Anki, and SuperMemo — the project that invented spaced repetition — opened a public API exposing its new SM-20 algorithm. Both moves point the same way: scheduling is no longer a fixed formula but a machine-learned model tuned to your individual memory.

If you build vocabulary from reading, this is good news worth understanding — but not for the reason the benchmarks suggest. The algorithm decides when a word comes back; your reading decides which words you collect and how well they stick. This guide explains what actually changed in 2026, how FSRS-6 and SM-20 differ, and where a smarter scheduler helps a reading-first learner versus where it is a distraction from the habit that really matters.

What changed in spaced repetition in 2026

The headline development is that the two dominant algorithm families both shipped machine-learned schedulers and made them widely accessible. FSRS-6 landed in late 2025, trained on roughly 700 million reviews contributed by about 20,000 volunteer Anki users, and from Anki 23.12 onward it sits in the scheduler with newer installs defaulting to it. The official Anki FAQ now documents FSRS as the recommended choice over the legacy SM-2 formula for almost all users.

On the commercial side, SuperMemo launched a public API on 31 March 2026 exposing its SM-20 algorithm — the first version in which every parameter is computed by machine learning rather than hand-tuned heuristics — with a free early-access tier capped at 100 repetitions a day. The significance is less about any single app and more about direction: as one 2026 survey of the field puts it, scheduling has moved from fixed formulas toward models that adapt to individual user data. Whatever tool you use, the engine underneath is now personalized in a way it simply was not a few years ago.

How FSRS-6 actually works

FSRS rests on a model of memory with three moving parts for every card. Stability is how durable the memory currently is — roughly, how long until you would forget it. Difficulty is how hard that particular card is for you. Retrievability is the probability you could recall it right now, which decays over time. The algorithm uses seventeen trainable weights to predict these values and schedules each card so it returns when your retrievability drops to a target you choose, typically 90%.

The reason this beats the old SM-2 formula is personalization. SM-2 applied the same interval multipliers to everyone regardless of how their memory actually behaved. FSRS optimizes its weights against your own review log, so a deck you find easy gets stretched intervals while a genuinely hard deck gets tighter ones. You do not need to understand the math to benefit — the one habit that matters is reviewing daily so the model has fresh data, and re-running the optimizer occasionally so the weights stay tuned to you. The broader science of why spaced, level-matched exposure works is covered in our piece on the research behind reading at your level.

FSRS-6 vs SuperMemo SM-20: which should you use

For the overwhelming majority of self-directed learners, the answer is FSRS-6, and the reason is access. FSRS is free, open-source, and already built into Anki and a growing list of reading apps, so you get a state-of-the-art scheduler with no setup beyond flipping it on. It has been validated on hundreds of millions of real reviews, and because it reads your existing history it improves the moment you switch an established deck to it. There is no subscription, no API key, and no per-day cap to think about.

SuperMemo's SM-20 is a serious algorithm with a long pedigree, and the new API is genuinely useful — but mostly to developers building scheduling into their own products, not to learners managing a deck. The free tier's 100 reviews-per-day limit and one-time import of 10,000 historical repetitions make it a building block rather than a daily driver. Unless you specifically live inside the SuperMemo ecosystem, you lose nothing by using FSRS. The meaningful takeaway is not which one wins but that both have abandoned fixed formulas, so the floor for everyone has risen.

Why the algorithm is the smaller lever for readers

Here is the part the benchmark wars tend to bury. A scheduler can only ever optimize the timing of a review; it cannot rescue a card that was badly made in the first place. A bare word-to-definition card lifted from a frequency list has one weak retrieval cue, so it lapses often no matter how cleverly it is scheduled — and every lapse pulls the interval back to the start. The biggest gains in retention come from the card itself, and that is decided before any algorithm gets involved.

Words mined from reading start with a structural advantage: they arrive wrapped in a sentence, a grammar pattern, and a situation you cared about understanding. That context gives your brain multiple hooks to retrieve the word later, which means fewer lapses and longer intervals from the very first review. Put a modern scheduler on top of context-rich cards and the two effects compound. This is exactly the read-capture-review loop we lay out in build vocabulary from reading — FSRS-6 simply makes that loop cheaper to maintain.

How to set it up without breaking your habit

If you use Anki, the whole upgrade is a few clicks: open the deck options, enable FSRS, set a desired retention of 90% to start, and review normally for a couple of weeks. After you have accumulated some history, run the FSRS optimizer so the weights fit your memory rather than the global defaults. Resist the urge to chase a higher retention target like 97% — it sounds appealing but balloons your daily review count for marginal gains, which is the fastest route to a backlog you start avoiding.

If you use a reading-first app that mines words for you, check whether it already runs FSRS under the hood — many of the 2026 generation do — and if so you get the benefit automatically while keeping cards tied to the sentences they came from. Either way, the non-negotiable habit is the same as it has always been: cap new cards at five to ten a day, review every day so the algorithm has fresh data, and feed the deck from content you actually read. The scheduler is the engine; your reading is the fuel.

Putting it together

The 2026 algorithm news is real and worth acting on: switch your decks to FSRS, set a sane 90% retention target, run the optimizer, and reclaim the 20–30% of reviews you were doing for no extra retention. SuperMemo's SM-20 API confirms the industry direction but is not something most learners need to touch. Treat the algorithm as a free efficiency upgrade, not a magic bullet.

Then spend the time you save where it compounds — reading. A better scheduler on top of words mined from real, level-matched content is the combination that actually moves your vocabulary, because reading supplies the context and the motivation that no algorithm can manufacture. To put the loop to work in a specific language, see our reading-first roadmaps for Japanese, Korean and Portuguese.