You may recognize this pattern: an answer feels obvious immediately after you read it, yet a few days later it is hard to recall. Meanwhile, familiar material keeps coming back. FSRS addresses the scheduling question between these experiences: when should this particular card return?
What is FSRS?
FSRS stands for Free Spaced Repetition Scheduler. It is an open-source algorithm for scheduling spaced repetition. Review history, elapsed time and your ratings inform its estimate of a card’s memory state and the next review date. It replaces the need to keep a separate calendar for every card.
Scheduling has a specific meaning: deciding when to ask again. It does not read an article for you, judge an unrecorded spoken answer, or certify mastery of a subject after one rating. Accurate content, clear questions and an honest attempt to recall still matter.
What happens during one review?
- Read the question and try to answer from memory before revealing the other side.
- Reveal the answer and compare it with what you actually recalled. Familiarity after revealing is different from recall beforehand.
- Rate this attempt. The rating describes this recall, rather than assigning you a permanent ability score.
- The system records the review and updates this card’s memory state and due date. Other cards retain their own histories.
In Deckumi’s current Web review implementation, four buttons map to the four FSRS ratings. The minutes or days shown below them preview the next interval for this card’s present state; they are not a fixed timetable shared by every card. After submission, the saved server due date is authoritative.
How should you choose a rating?
Rating | What happened this time | How to decide |
|---|---|---|
Again | You could not recall the answer, or missed a key part. | Choose this even if the answer feels familiar after revealing it. |
Hard | You recalled it correctly, but with considerable effort. | You must have answered correctly before revealing. |
Good | You recalled the correct answer normally. | A suitable rating for an ordinary successful recall. |
Easy | You recalled it quickly and comfortably. | Little searching was needed, without an answer hint. |
The most important distinction is Hard versus Again. Hard means difficult but successful recall. It does not mean that you forgot but would rather avoid a failing grade. Repeatedly using Hard after forgetting sends an overly optimistic signal. Accurate feedback is more useful than impressive-looking ratings. See the official rating guidance.
For example, a card asks which form of energy photosynthesis converts light into. If you answered chemical energy after a long effort, Hard may fit. If you only recognized the answer after revealing it, use Again. When two ratings seem similar, first decide whether recall succeeded, then judge the effort.
Three ideas: difficulty, stability and retrievability
Difficulty D describes how difficult a card is to retain. Stability S describes resistance to forgetting. Retrievability R estimates the probability of recalling it at a particular moment. These serve different purposes: elapsed time changes the estimate of recall now, while a new review updates the memory estimate.
FSRS expresses stability in days, using the interval at which estimated retrievability reaches 90% as its reference. Stability is not a percentage of mastery or a rule forbidding earlier study. Implementations evolve across versions; understanding these concepts is more useful than memorizing one version’s formula.
See the FSRS algorithm documentation for terminology and mathematical definitions. Here they explain the interface, rather than guarantee an interval or outcome for every learner.
Does 90% desired retention mean a 90% exam score?
No. Desired retention is a scheduling target for recall when cards become due. It is not an exam score, a measure of understanding, or a promise that exactly nine out of ten cards will be answered correctly. Actual reviews are needed to check how well predictions fit experience.
A higher target generally means shorter intervals and more reviews. A lower target generally reduces the workload while accepting more forgetting. It is a trade-off between retention and study effort, rather than a number that is always better when increased.
Deckumi’s current Web scheduling policy uses a target of 0.9 and a fixed, versioned parameter profile. It does not expose personal parameter training or a target-retention setting. Options available in another application are not necessarily Deckumi features. See the Anki manual for further discussion of desired retention.
Why can one card return soon and another much later?
New cards need an initial learning process, and forgotten cards may enter relearning. Deckumi’s current Web policy includes minute-scale learning and relearning steps, so a recently seen card may become due again soon. Longer-term intervals are calculated from the memory state and rating.
The same Good rating can produce different intervals. Two cards may have different histories, and one card’s state changes over time. If you return several days late, the actual elapsed time matters; a scheduler should not pretend the review happened on its original date.
An all-caught-up message means there are no due or new cards waiting right now. It does not mean that every card is permanently mastered. Cards can enter the queue again when they become due.
What do Deckumi’s five memory-strength labels mean?
The four ratings are your input. The five memory-strength labels are Deckumi’s simplified display of saved state. They are different systems; there is no fifth FSRS mastery rating. The current interface groups stability S as follows:
Label | Current display rule | Meaning |
|---|---|---|
Not studied | No saved FSRS memory state | No memory record has been established yet. |
Starting | S < 1 day | Estimated stability is still low. |
Building | 1 day ≤ S < 7 days | Stability is building. |
Stable | 7 days ≤ S < 30 days | The stability estimate is longer. |
Long-term | S ≥ 30 days | Estimated stability is higher; due reviews still matter. |
These thresholds are Deckumi interface rules, not five official FSRS levels or independently validated competence bands. They should not be converted into 20%, 40% or 100% mastery. Long-term does not mean that forgetting is impossible.
Clear cards matter more than flattering ratings
If one card asks for a definition, formula, example and exception together, a partial answer is hard to rate. Splitting it into focused questions can make the criterion clearer. If an answer is ambiguous or wrong, correct the content before continuing.
For example, “Explain all of photosynthesis” gives an unclear boundary for success. “Which energy form does photosynthesis convert light into?” has a clearer target. A separate why-question can practice explanation and application, rather than treating one memorized term as the whole subject.
How does Deckumi currently keep these records?
The current Web implementation keeps a memory state and due date per card. Signed-in reviews are saved through the server; guest practice remains in the current browser. This does not imply that native apps already use the same cloud review flow or that unreleased cross-device features are available.
Changing a note, image size, position or caption font size does not clear saved memory state. Changing the question, answer, or the meaning carried by an image or caption does require a new memory state for that content. Familiarity with the old question should not be assigned automatically to a new one.
This implementation uses the open-source ts-fsrs scheduler. Deckumi behavior described here is based on the Web implementation as of 21 October 2026; the article explains the learning flow and does not assert deployment to every release channel.
Three things to remember when starting
- Try to recall before revealing. Rate what actually happened.
- Use Again when you forget. Hard is successful but effortful recall, not a gentler word for forgetting.
- Keep returning for due reviews and improve ambiguous cards. Memory estimates help arrange practice; they do not replace understanding or application.