How to Optimize Spaced Repetition (FSRS Retention Guide)

The short answer

Set your FSRS desired retention between 0.85 and 0.95. For most learners, 0.90 (90%) is the right starting point: it keeps daily review counts manageable while ensuring you actually remember what you study. Going higher (say, 0.95) dramatically increases review volume for a small memory gain. Going lower (0.80 or below) saves time per day but causes so many lapses that you end up relearning cards anyway, often increasing total workload.¹

Beyond picking a number, optimizing spaced repetition means letting FSRS auto-tune its parameters to your memory, grading honestly, keeping new-card volume sustainable, and reviewing consistently. This article covers what the retention target actually controls, how to choose your value, and the habits that let FSRS do its best work.

---

What "desired retention" means in FSRS

In FSRS, desired retention (DR) is the probability of recall you want when a card comes due. If you set DR to 0.90, FSRS schedules each card so that, at the moment it appears in your queue, you have roughly a 90% chance of remembering it. The algorithm achieves this by adjusting intervals: higher DR produces shorter intervals (more frequent reviews), and lower DR produces longer intervals (fewer reviews, more forgetting).

Under the hood, FSRS uses the DSR model (Difficulty, Stability, Retrievability) to track your memory state for every card.² Retrievability (R) is your current probability of recall, decaying exponentially over time along the forgetting curve first documented by Ebbinghaus and replicated in modern studies.³ Stability (S) is the time in days for R to drop from 100% to 90%. Difficulty (D) captures how inherently hard the material is. When R drops to your desired retention value, FSRS schedules the review. For a full explanation of the DSR model, see how spaced repetition works.

How to pick the right retention target

The 0.85 to 0.95 sweet spot

FSRS allows values from 0.70 to 0.99, but the practical range is narrower. Below 0.85, you forget so many cards that relearning costs outweigh the shorter intervals. Above 0.95, intervals shrink so aggressively that your daily review count can double or triple for only a few percentage points of extra retention. The workload curve is U-shaped: there is a minimum somewhere in the middle, and it varies by learner.¹

• 0.90 (default): The recommended starting point. Balanced workload and strong recall. Most users should begin here.
• 0.85 to 0.89: Good for large decks or time-limited learners who can tolerate a few more lapses in exchange for fewer daily reviews.
• 0.91 to 0.95: Suited for high-stakes material (medical exams, professional certifications) where forgetting even a small fraction is costly. Expect noticeably more reviews per day.

Use the "compute minimum recommended retention" feature

FSRS includes an optimizer that analyzes your review history and calculates the retention value that minimizes your workload relative to knowledge gained. In Anki 24.04 and later, you can click "Compute minimum recommended retention" after optimizing your parameters. This gives you a data-driven starting point personalized to your memory, not a guess.

How FSRS auto-tunes its parameters

Unlike older algorithms such as SM-2, FSRS does not rely on fixed multipliers. It uses Maximum Likelihood Estimation to fit 19 internal parameters (as of FSRS v5) to your actual review history.² The optimizer compares its predicted recall probabilities against your real outcomes. If FSRS predicted you would remember 90% of a batch and you actually remembered 95%, it adjusts its weights so future predictions are more accurate.

This self-correction makes FSRS fundamentally different from heuristic schedulers. SM-2 uses static ease factors that drift over time, a problem known as "ease hell." FSRS retrains on your data, so its model of your memory improves with every optimization cycle. You can re-optimize periodically (every few months or after a few thousand reviews) to keep the model current.

What desired retention actually affects

Review load vs. forgetting

The relationship between desired retention and daily workload is not linear. Raising DR from 0.85 to 0.90 adds a moderate number of reviews. Raising it from 0.90 to 0.95 can nearly double your load, because intervals shrink exponentially as you approach perfect retention. Meanwhile, the actual knowledge gain from 0.90 to 0.95 is relatively small in absolute terms.¹

On the other end, dropping DR below 0.80 backfires. You lapse so often that the algorithm must reset cards to short intervals, and the relearning cost erases most of the time you saved. Research on spacing confirms that an optimal gap exists for any given retention interval: too-short gaps waste effort, and too-long gaps cause excessive forgetting.⁴⁵

Long-term memory formation

Each successful retrieval at a low retrievability point strengthens stability more than an easy recall at high retrievability.² This is the "desirable difficulty" principle in practice. A well-chosen retention target keeps reviews just hard enough to maximize memory strengthening per session, without pushing you into failure spirals that demoralize and waste time.

Practical tips for optimizing your spaced repetition

1. Review consistently

FSRS assumes you review cards around the time they are due. Skipping days causes retrievability to drop well below your target, which means more lapses and more relearning. Consistent daily sessions, even short ones, keep the scheduler's predictions accurate and your workload stable. Cramming a week's backlog in one sitting is massed practice, not spaced repetition.⁴

2. Grade honestly

FSRS learns from your grades. If you press "Good" when you actually guessed or peeked, the algorithm overestimates your stability and stretches intervals too far. You will fail the card later and wonder why it is not working. Grade based on genuine recall: if you needed a hint or hesitated significantly, that is "Hard" or "Again." Honest data is the single most important input you can give any SRS algorithm. See common spaced repetition mistakes for more on this.

3. Do not overload new cards

Every new card you introduce today becomes a stream of future reviews. Adding 30 or 50 new cards per day feels productive in the moment, but within weeks your due queue explodes. Set your daily new-card limit based on how many total reviews (new plus due) you can sustain, not on how many fresh items you want to see. When your backlog grows, pause new cards until dues stabilize.

4. Optimize parameters periodically

Re-run the FSRS optimizer every few months or after accumulating a significant number of new reviews. Your memory characteristics shift as you learn, and the algorithm cannot adapt if it never sees updated data. In apps that support FSRS, this is typically a one-click operation under deck or scheduling settings.

5. Match retention to your goals

Not every deck needs the same desired retention. Vocabulary for daily conversation might work well at 0.88. Medical terminology for board exams might warrant 0.93. FSRS supports per-preset retention targets, so you can tailor the trade-off for each subject rather than forcing one number across everything.

How LinGoat takes care of all of this for you

LinGoat is designed so you never have to think about any of the tuning above. Every optimization lever described in this article is handled automatically:

• FSRS runs under the hood. Every word and grammar point you practice is tracked with the full DSR model (Difficulty, Stability, Retrievability). LinGoat sets and maintains your retention target so you get strong recall without drowning in reviews.
• Honest grading is built in. Because LinGoat uses production-based practice (you write full sentences, not tap multiple-choice answers), the recall signal is clean: either you produced the word correctly or you did not. There is no way to accidentally inflate your scores by guessing from a list of options.
• New-card flow is automatic. LinGoat controls how many new words enter your study queue based on your current review load and pace. You never manually set a "new cards per day" limit or worry about overwhelming the schedule. Due reviews always take priority over fresh introductions.
• Laddering solves the first-turn problem. Before a word ever enters the main FSRS schedule, it goes through LinGoat's dedicated laddering system: priming (you type the word while it is visible), a short interleaved distractor that clears working memory, and blind active recall. This ensures the first data point FSRS sees reflects real memory, not a lucky guess from a sensory echo. Without this step, early scheduling data is noise. Read the full design in Fixing the First-Turn Bottleneck, or see why you should not put new words straight into SRS for the science behind it.

The result: you sit down, practice sentences in your target language, and LinGoat handles retention tuning, grading integrity, card pacing, and first-contact encoding in the background. See how LinGoat works or start practicing to try it.

References

1. Ye, J., Su, J., & Cao, Y. (2022). A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 4381-4390. https://doi.org/10.1145/3534678.3539081
2. Su, J., Ye, J., Nie, L., Cao, Y., & Chen, Y. (2023). Optimizing Spaced Repetition Schedule by Capturing the Dynamics of Memory. IEEE Transactions on Knowledge and Data Engineering, 35(10), 10085-10097. https://doi.org/10.1109/TKDE.2023.3251721
3. Murre, J. M. J., & Dros, J. (2015). Replication and Analysis of Ebbinghaus' Forgetting Curve. PLOS ONE, 10(7), e0120644. https://doi.org/10.1371/journal.pone.0120644
4. Carpenter, S. K., Pan, S. C., & Butler, A. C. (2022). The science of effective learning with spacing and retrieval practice. Nature Reviews Psychology, 1, 496-511. https://doi.org/10.1038/s44159-022-00089-1
5. Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., & Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11), 1095-1102. https://doi.org/10.1111/j.1467-9280.2008.02209.x