FSRS: Stability when first non-ignored revlog has a relearning interval

mbrubeck · April 28, 2025, 12:51am

This is a possible bug that exists in both Anki 25.02.x and Anki 25.05 (Beta 1 and 2), and possibly earlier versions too.

The bug happens when ignore_revlogs_before is set, and a card’s first review after the ignore date is an “Again” review that sets the interval to a relearning interval. For example, this card, in a deck where the “Ignore cards reviewed before” cutoff is in 2024:

Because my relearning interval is the default 10 minutes, this causes fsrs_item_for_memory_state to use its minimum interval of 1 day when calling memory_state_from_sm2:

github.com/ankitects/anki

rslib/src/scheduler/fsrs/memory_state.rs

6ff023f6a


      
          let mut item = output.fsrs_items.pop().unwrap().1;
          if output.revlogs_complete {
              Ok(Some(FsrsItemForMemoryState {
                  item,
                  starting_state: None,
                  filtered_revlogs: output.filtered_revlogs,
              }))
          } else if let Some(first_user_grade) = output.filtered_revlogs.first() {
              // the revlog has been truncated, but not fully
              let first_review = FirstReview {
                  interval: first_user_grade.interval.max(1) as f32,
                  ease_factor: if first_user_grade.ease_factor == 0 {
                      2500
                  } else {
                      first_user_grade.ease_factor
                  } as f32
                      / 1000.0,
              };
              let mut starting_state = fsrs.memory_state_from_sm2(
                  first_review.ease_factor,
                  first_review.interval,

I believe that when the first non-ignored revlog has a rating of 1 (Again) and is followed by one or more Relearning revlogs, the last (“graduating”) Relearning revlog should be used rather than the initial “Again” revlog. In this case, that would result in an initial interval of 2.5 months rather than 1 day.

mbrubeck · April 28, 2025, 4:20am

I submitted a pull request with a possible fix for this: Don't use relearning step as starting memory state by mbrubeck · Pull Request #3959 · ankitects/anki · GitHub

Topic		Replies	Views
Bug: Bad reschedule behavior when "ignore reviews before certain date" is on FSRS	11	96	March 31, 2025
[FSRS] failing cards that have no review history does not decrease their intervals Help	18	893	December 20, 2023
Interday intervals for lapsed cards results in disproportionately short intervals FSRS	3	57	March 22, 2025
"reset" card without FSRS optimizer ignoring previous reviews FSRS	4	132	August 21, 2024
Allow to not move card to learning with empty relearning steps Suggestions	3	55	December 15, 2024

FSRS: Stability when first non-ignored revlog has a relearning interval

Related topics