Does
mw.col._set_enable_fsrs_short_term_with_steps(False)
disable it
Looks better now:
1 Like
Hmm but why cant it go lower than 9.22 minutes 
Do you know the reason behind this
Because the minimum value of stability is 0.01. And 0.01 day = 864 seconds = 14.4 minutes. Your desired retention is 95%, so the interval is shortened from 14.4 minutes to 9.22 minutes.
6 Likes
This makes sense. I wish that in the future FSRS would be able to calculate its intervals based on the existing custom intervals 
But this is surely a good fix. Thanks! 
I know this may come off as dumb, as I am aware this has been explained before, but why is it that 0.01 is the minimum value of stability, or why that there should be a fixed minimum in the first place
Because if the stability is too small, it will take a lot of reviews to increase it. In previous version, FSRS doesn’t consider short-term reviews, so it’s worse.
2 Likes
Will a future version allow for a lower minimum stability value for short-term reviews?
2 Likes
If I figure out the short-term memory model, I will decrease the minimum stability limit in the same version.
7 Likes
Does it mean, that short-time FSRS will never put a step smaller than 14.4 minutes? (with 0.9 target retention set)
It can be a little dangerous. When someone goes through new cards, with a lot of them failing, it could accumulate quite a backlog within 14 minutes. The default first step of 1 minute is a remedy, to not make such an accumulation.
Would a 1 minute interval not be somewhat counterproductive for recall?
From what I remember in the papers I’ve read, optimal initial short-term intervals after encoding typically range from 15 to 30 seconds. I also recall that a certain degree of difficulty is necessary to retain the memory trace.
For instance from Implications of Short Term Memory Research for the Design of Spaced Repetition Based Mobile Learning Games | IEEE Conference Publication | IEEE Xplore :
The main difference between the two memory stores [short term vs long term] is the time span information can reside in them. Information usually remains in working memory for about 20 to 30 seconds unless it is prevented from disappearing by rehearsing [which references https://labs.la.utexas.edu/gilden/files/2016/03/1545ca4cdad4f480f2fab8fd505a5b7f7603.pdf].
[…] forgetting occurs as a result of the automatic decay of the memory trace, which happens in short term memory as early as 15 to 30 seconds after learning the information unless it is rehearsed [which references Decay Theory of Immediate Memory: From Brown (to Today (2014) - PMC].
[…] in order to achieve the spacing benefit, there has to be some level of retrieval difficulty during rehearsals. This can be accomplished by certain spacing intervals between the repetitions. If this is not the case, for example when engaging in massing, the content might be successfully retrieved, but the memory trace will not be strengthened since the similarity between the first encoding session and the rehearsal session is too high. This results in little difficulty to retrieve the content from long term memory. In this case, there is only little reconsolidation of the memory trace. On the other hand, if certain time passes between the rehearsal sessions, the familiarity with the learning content declines and a greater strengthening of the memory trace occurs. This is also supported by the effort a learner needs to take, which leads to a certain modification of the memory trace when doing a rehearsal. However, if the interval between the learning sessions is too long, the content can be forgotten and the learning process might need to be repeated from the beginning [which references APA PsycNet, APA PsycNet and APA PsycNet].
2 Likes
I suppose it begs the question, then: if short-term encoding was the goal, or beneficial, why do we not see recommendations or default settings for anything near that value?
1 Like
The values are for after encoding.
Other than that, I have no idea to be honest. Maybe it’s challenging to implement, or there may not be sufficient evidence or testing to support those values. Or there could be other factors that render this particular factor negligible or complicate its integration.
1 Like
To add more: the default 1 minute learning interval is here because there is a work saying, that a person can keep usually up to 7 pieces of information in their working memory at once. This was confirmed by Dae in one of the topics.
With 14 minutes delay, one can accumulate almost 100 cards. Then it will be painful and counter-productive to circle through 100 unknown cards until you learn them - 1 minute is to create smaller learning batches for a user.
3 Likes
Creating a backlog of 100 cards in 14 minutes seems seems a bit of an extreme case, to my mind. Under what conditions is someone getting a card wrong every 8 seconds for 100 consecutive cards? Or, rather, at that point I would think the issue is potentially not the interval/s, necessarily
Perhaps I’m mistaken, but the 7 pieces idea refers to working memory which is not typically the type of memory or recall for which Anki is designed, no? That’s not to say that it’s not at all involved, but I find it hard to shake the idea that a 1 minute interval is in all likelihood an unadvisedly short interval.
Of course there are boundary cases, and extremes, but at the very least surely there are diminishing returns with an interval of 60 seconds. At that point my suspicion would be that other issues are likely afoot if one needs to see a significant number of cards with that small of an interval.
Open to hearing arguments to the contrary, but I I think I remain with some respectful skepticism of the utility/wisdom of a 1-minute interval.
2 Likes
I am not saying, that as low as 1 minute is necessary. However 14.4 minutes as a minimum seems quite high in some cases.
2 Likes
@L.M.Sherlock has said that it is not that he is not willing to lower the minimum stability (and thereby lowering the 14.4 minimum interval). He just couldn’t figure out a model that would allow him to do so. If he could, he would.
The question is not about the IF, but about the HOW?
1 Like
Have you had a chance to read the studies, particularly the cited parts? They suggest sub-minute intervals after encoding. Though I’m uncertain whether these intervals apply in practice. At least the way I study, I tend to retrieve information over the span of many minutes after encoding before creating cards. It’s possible these findings are more relevant to pre-made decks, where both encoding and initial memorization take place through Anki, or there may be instances where some cards haven’t undergone either the encoding phase or initial memorization. It might be we even agree after all, albeit for different reasons. Let’s see if L.M.Sherlock finds a solid model for short term memory, as that will put to rest our doubts.
1 Like
Crude idea, but should be easy to implement: allow stability to go as low as 0.01. But after a Good or Easy answer stability should be no less than 0.1.
This will allow small intervals (less than 1.5 minutes) if a card is answered Again many times, but will bump up recovery after the correct answer.
@L.M.Sherlock What do you think?
5 Likes