Findings For newly introduced cards, the average intraday retention corresponding to different initial ratings are: Again: 74.2% Hard: 91.7% Good: 96.2% Easy: N/A (the card is graduated immediately) For lapsed cards, the average intraday retention is 85.3%. Data and code are available at: Anki-button-usage/analysis.ipynb at main · open-spaced-repetition/Anki-button-usage · GitHub The current default learning steps are set at 1m 10m, with the default relearning step at 10m. This translates to the following intervals for new cards based on ratings: Again: Review after 1 minute Hard: Review after 6 minutes Good: Review after 10 minutes Lapsed cards are reviewed after 10 minutes. Suggestion For typical users: New cards: The ‘Again’ interval is excessively long The ‘Good’ review interval is insufficiently long Lapsed cards: The interval is excessively long To elevate the intraday retention from 74.2% to 90%, the interval should be reduced by a factor of 3.5. Thus, 60s / 3.5 ≈ 17s. To decrease the intraday retention from 96.2% to 90%, the interval should be increased by a factor of 3. Hence, 10m * 3 = 30m. To raise the intraday retention for lapsed cards from 85.3% to 90%, the interval should be reduced by a factor of 1.6. Therefore, 10m / 1.6 ≈ 6.25m. [image] [image] [image] Given these calculations, the proposed adjustments are: Default learning steps: 17s 30m Default relearning step: 6m For a more moderate approach, consider 30s 15m as the default learning steps.

[image] L.M.Sherlock: Expt/step optimizer (WIP) by L-M-Sherlock · Pull Request #487 · open-spaced-repetition/fsrs4anki-helper · GitHub Now it works for separate decks, thank you!! Here is the table for one deck. [image] Here is the table for another deck(I have never used the hard rating here). [image] But how can you interpret this data here :question: For example, what does Review Timing Distribution mean :question: And could this eventually be used to allow the auto short-term scheduler to take custom learning steps into account :question: [image] FSRS 5: <1d Scheduling and Learning Steps FSRS FSRS-5 doesn’t consider the real interval of your learning steps because I haven’t figured out how to model the short-term memory accurately. So only the ratings matters in FSRS-5.

[image] anon12869913: Is there enough information in these stats shown to give the correct relearning step(s)? [image] I will update the PR soon. Edited: Comment by [L-M-Sherlock] L-M-Sherlock - Expt/step optimizer (WIP) open-spaced-repetition:main ← open-spaced-repetition:Expt/step-optimizer Updated: Patch: [fsrs4anki-helper.ankiaddon.zip](https://github.com/user-attachments/files/17686974/fsrs4anki-helper.ankiaddon.zip)

I still don’t understand what these numbers entail. And how I should deal with them.

You can read this wiki page: Quartile - Wikipedia [image] The first row shows when I press again in a new card, 25% of the next reviews are done in 1.2 min, 50% are done in 1.38 min, 75% are done in 2.22 min. The average retention of the next review is 83.48%. And the estimated stability is 0.8 min.

Amazing! This seems a really good step forwards for automated intraday intervals management. Does it make sense for something based on these statistics to eventually replace the current optimizer for initial FSRS stabilities (w0-w3 iirc)?

[image] anon12869913: Does it make sense for something based on these statistics to eventually replace the current optimizer for initial FSRS stabilities (w0-w3 iirc)? The challenge is how to build a bridge between short-term memory and long-term memory.

Ohhhh. Now I starting to see things a bit more. So this would change based on whatever learning steps I have, and based on that, the stability is calculated. If I remember correctly, Stability shows the time after which the retrievability has declined to 90%, so if I want a retention of 95%, I should aim for less time. My learning steps for this is 1s 3m (auto intraday scheduler turned on).So what would be the recommendation here.:question: Or does the short-term scheduler not yet have a link to custom learning steps :question: [image] [image] bruts: L.M.Sherlock provided a better answer. As for how to deal with them, I suppose you have to wait less between a review and the next, thus adjusting the stability to desired retention should help. It should bring your delays closer to current 25th percentile, so for example, reducing your delay from 29 seconds to something closer to 9 seconds. Needs to be seen if it works in practice. So If I understand this correctly: 1-My learning steps are 1s 3m. So I have a 1s again interval (0.02 min) 2-Having a 1s again interval does not necessarily mean I would see the card in 1s → delays (which is true from personal experience) 3-Having this delay influences my retention obviously 4-Based on the delays and retention achieved, stability is calculated. 5-In my case stability is 0.02 min, so I should be indeed be seeing the card in under 0.02 if I want a retention of anything >90%. If I did understand this correctly, my question now is: would the stability calculated still be the same if I had different learning steps and in turn different delays :question:

The stability is calculated from your review data. And changing the learning steps will let you generate different review data in the future. It may affect the stability. However, in theory, the stability only depends on your first rating unless you change your rating habit.

By the way, I tend to add more detail stats here. In your opinion, is it too verbose? [image]

I liked it before. T25% won’t be easy for non-math people.

Suggestion about the learning steps after analyzing the Anki Dataset

Anki Scheduling

L.M.Sherlock November 9, 2024, 12:04pm 48

The challenge is how to build a bridge between short-term memory and long-term memory.

FSRS 5 shortest interval

Topic		Replies	Views
FSRS 5: <1d Scheduling and Learning Steps FSRS	167	4650	November 25, 2024
Several FSRS-related suggestions FSRS	167	3826	October 29, 2024
Clarify what optimal retention means FSRS	77	4313	January 9, 2025
[Feature Request] FSRS Should Ignore Learning Cards Reviews FSRS	36	1167	June 29, 2025
[FSRS] Separate Desired Retention for Young and Mature cards FSRS	32	1354	July 18, 2026

Suggestion about the learning steps after analyzing the Anki Dataset

Related topics