Definition:
Retention rate is the percentage of flashcard reviews in which a learner successfully recalls the target information. An 85% retention rate means the learner answers correctly on 85 out of every 100 review attempts. It is one of the most important metrics in spaced repetition systems (SRS) because it directly reflects how well learned material is being retained in long-term memory.
How Retention Rate Is Calculated
At its simplest:
> Retention Rate = (Correct Reviews ÷ Total Reviews) × 100
In Anki, this is reported in the Statistics panel. FSRS tracks true retention — the probability that a card is recalled at the moment it is due, not just on average across all reviews. This distinction matters because cards due after a long interval are harder than freshly reviewed ones.
What Is a “Good” Retention Rate?
There is no universal ideal, but common target ranges are:
| Target Retention | Tradeoff |
|---|---|
| 70–79% | Fewer reviews, more forgetting |
| 80–90% | Standard recommended range |
| 91–95% | Diminishing returns; review burden rises steeply |
| 96–99% | Very high workload; rarely practical for large decks |
The SM-2 algorithm historically targeted around 90% retention. FSRS allows users to set a custom target retention (default: 90%) and optimizes card scheduling to hit that number efficiently.
The relationship between retention rate and review load is not linear — pushing from 90% to 95% roughly doubles review frequency for the same deck.
Why Retention Rate Matters
- Efficiency signal: A rate consistently below 70% may mean cards are too hard, too infrequently reviewed, or poorly written.
- Progress tracking: Stable high retention over weeks indicates that material is genuinely entering long-term memory.
- Optimizer input: FSRS uses observed retention data to calibrate memory parameters (stability, difficulty) per card.
- Learning vs. reviewing balance: Low retention usually means more time re-learning instead of acquiring new material.
Factors That Affect Retention Rate
- Interval length: Longer intervals naturally produce lower retention — this is the Forgetting Curve in action.
- Card quality: Vague, complex, or poorly contextualized cards are harder to retain than clear, atomic ones.
- Ease Factor / Difficulty: In FSRS, each card has a difficulty score; harder cards are scheduled more frequently to compensate.
- Learning stage: Newly introduced cards naturally have lower stability and will be forgotten more often until they consolidate.
- Sleep and health: Memory consolidation is heavily tied to sleep; poor rest reduces retention rates across all decks.
Retention Rate in FSRS vs. SM-2
| Feature | SM-2 | FSRS |
|---|---|---|
| Target retention | ~90% (implicit) | User-defined (default 90%) |
| How it’s tracked | Through Ease Factor adjustments | Predicted probability of recall |
| Optimization | Manual ease adjustments | Full parameter optimization via review history |
| Accuracy | Approximate | More precise; better handles lapses and long gaps |
FSRS’s concept of true retention (the probability of recall at due date) is a more rigorous measure than simply counting pass/fail across all cards regardless of interval.
History
The concept of retention rate in spaced repetition systems traces to Hermann Ebbinghaus’s (1885) experimental research on memory, which established the forgetting curve — the exponential decay of memory over time without review. Ebbinghaus’s work showed that retention could be measured as a percentage of material successfully recalled at specific intervals. Pimsleur (1967) applied retention rate concepts to language learning through graduated interval recall. Wozniak’s SuperMemo (1987) formalized retention rate as a target parameter in SRS algorithms: users set a desired retention rate (e.g., 90%) and the algorithm schedules reviews to maintain that target. Modern algorithms like FSRS (2022) made retention rate a directly adjustable parameter, allowing learners to set their target and have the algorithm optimize review scheduling accordingly.
Common Misconceptions
“A higher retention rate is always better.”
Targeting 99% retention rate requires extremely frequent reviews — dramatically increasing daily review load while providing diminishing returns. The 85-90% range represents the empirically supported sweet spot: high enough to maintain knowledge, low enough to avoid review overload.
“Retention rate measures how well you know the material.”
Retention rate measures the percentage of successful recalls during review sessions — it reflects testing performance under SRS conditions, not necessarily the ability to use the material in real communication. A 90% retention rate on vocabulary cards doesn’t guarantee 90% recall in conversation.
“Low retention rate means the SRS isn’t working.”
A retention rate below target may indicate that intervals need adjustment, card quality needs improvement, or the material is too difficult — not that spaced repetition itself is failing. It’s diagnostic information, not a system failure.
“Retention rate should be the same for all material.”
Different types of material (vocabulary recognition, production, grammar patterns, kanji readings) may naturally have different retention rates. Some learners set different targets for different decks rather than using a universal target.
Criticisms
The retention rate metric has been criticized for encouraging a narrow focus on measurable recall at the expense of deeper learning goals. Optimizing for retention rate on flashcard reviews may lead learners to spend excessive time on SRS review while neglecting input activities (reading, listening) that develop comprehension and fluency.
The metric itself is an oversimplification: a binary “recalled/not recalled” judgment doesn’t capture the nuances of word knowledge — partial knowledge, recognition without production ability, and contextual variation are all collapsed into pass/fail. Additionally, retention rate consistency varies across individuals and even across study sessions for the same individual — fatigue, attention, and mood all affect recall performance, meaning that the retention rate on any given day may not reflect stable knowledge. The FSRS algorithm addresses some of these issues through its probabilistic modeling, but the fundamental limitation of binary recall assessment remains.
Social Media Sentiment
Retention rate is actively discussed in Anki and SRS communities on Reddit and Discord. The optimal retention rate setting is one of the most common questions from new Anki users, with experienced users typically recommending 85-90% for vocabulary and slightly lower (80-85%) for less critical material. The introduction of FSRS made retention rate a directly adjustable parameter, generating extensive community discussion about optimal settings.
Common misconceptions in community discussions include the belief that higher retention rate is always better (leading to review overload) and the conflation of SRS retention rate with real-world recall ability. The community generally emphasizes that retention rate is a tool for managing review workload, not a measure of true language mastery — “90% retention in Anki doesn’t mean 90% recall in conversation” is a frequently repeated insight.
Practical Application
- Aim for 85–90%. This is the sweet spot for most learners — high enough to feel confident, low enough to keep review load manageable.
- Check your stats weekly. A sliding retention rate can indicate deck problems (cards too hard, intervals too long) before they become overwhelming.
- Don’t chase 99%. The extra reviews needed for near-perfect retention rarely justify the time cost.
- Review your leeches. Cards that are repeatedly failed drag down retention rate and signal a need for card redesign or mnemonic help.
Related Terms
Retention rate is closely tied to Spaced Repetition, the Forgetting Curve, Ease Factor, SM-2, and FSRS. It is a practical output of the theoretical model first described by Hermann Ebbinghaus and later operationalized by Piotr Wozniak in SuperMemo.
Research
- Wozniak, P. A., & Gorzelanczyk, E. J. (1994). Optimization of repetition spacing in the practice of learning. Acta Neurobiologiae Experimentalis, 54(1), 59–62.
- Settles, B., & Meeder, B. (2016). A trainable spaced repetition model for language learning. Proceedings of the 54th Annual Meeting of the ACL, 1848–1858.
- Ye, J. (2023). A stochastic shortest path algorithm for optimizing spaced repetition scheduling. FSRS4Anki documentation.