90% Actual Retention with 90% Desired lead to lower and lower stability

[sound]

Scenario : You have DR=90%, and your R is exactly 90%, with 9 good answers following by always a wrong one.

With FSRS, with the default params, it leads to diminishing stability :

Params :
0.1115, 0.3729, 2.2811, 28.1747, 7.1026, 0.8136, 2.1284, 0.0021, 1.0378, 0.1748, 0.6346, 1.8757, 0.0764, 0.2309, 2.2853, 0.0027, 4.9064, 0.2462, 0.7301

With mine, same :

0.1115, 0.3729, 2.2811, 28.1747, 7.1026, 0.8136, 2.1284, 0.0021, 1.0378, 0.1748, 0.6346, 1.8757, 0.0764, 0.2309, 2.2853, 0.0027, 4.9064, 0.2462, 0.7301

I invite anyone reading this to try the same here :
https://open-spaced-repetition.github.io/anki_fsrs_visualizer

As reviews, just pass this : “133333333313333333331333333333133333333313333333331333333333133333333313333333331333333333”

Some might say : Yeah but interval gets very large in the first example. Sure. But for example in mine, the stability goes max to 30d before dropping to 6d.

I’m curious what other people parameters might give them ?

[Danika_Dakika]

[You’ve got the same parameters in your post twice, which I don’t think you intended.]

I feel like the piece you’re skipping over is that you’re not just getting that card wrong 1 out of 10 times – you’re getting it wrong after a shorter and shorter interval in each of those cycles. And sometimes after an interval where you’ve gotten it right before. That’s showing FSRS that your intervals need to grow more slowly. I don’t think there’s anything unexpected about that.

The reason stability is dropping is that you’re applying an unnaturally consistent grading sequence that would be rare in actual use. You’re more likely to get an answer wrong based on the length of the interval, not because you’ve gotten it right the last 9 times. Exactly 9 answers + slower growing intervals = lower and lower S.

A truer test would be to have the failure point come when the card gets to a similar or slightly longer interval. You can still look at a 10% overall failure rate without it being evenly spaced. Something more like this –

The sequence I used – although yours may vary with your parameters – 133331333333133333333133333333313333333333133333333333133333333333133333333333133333333333