I have a special deck for cards I do not remember well. In the preset of that deck, I set the desired retention to 95%, so these cards get enhanced training. What happened until just recently is that the interval of such a card remained at 2 days for about 10 consecutive times, then went up to 4 for a number of turns, and slowly crept up. Once a card went over 10 days, I moved it back to its ordinary deck.
HOWEVER, today I optimized the preset (unfortunately without taking a back up first), and what happened is that it does not increase the interval from 2 to 4, it remains at 2. I donât know of course how long it will do that. Even cards that already had a 4 day interval are set to 2 now, as can be seen in this example
That means that FSRS, after optimizing it with more data, came to the conclusion that the intervals should be shorter than what you had in the past. Thatâs not a bad thing as FSRS determined that shorter intervals help you better to remember your material.
Iâm not sure but maybe automatic backups include the FSRS parameters. In that case, you could restore an automatic backup and get back the FSRS behavior that you prefer (and which FSRS considers worse). See Backups - Anki Manual.
Well, I do have an automatic backup, but itâs two days old. Rather than reverting I would rather go on with the optimized FSRS, but I feel a bit thrown backâŚ
Your automatic backup shouldnât be more than about 30-minutes-old, if youâve been using Anki in the past 2 days. If you havenât been using Anki in the past 2 days, then 2-days-old is probably recent enough.
Was it the first time you optimized that preset? Then you were using the default parameters before, and you can reset it to those (click the circle-arrow next to the field, and then confirm). But as Anon pointed out â those parameters are worse for you, based on your review history.
After you change the DR, the cards will be scheduled using that DR the next time you study them. But the card in your example has a stability of 1, so I wouldnât expect it to jump up to 4d any time soon.
You can use the FSRS Visualizer to see how long it will take your cards to advance with this new set of parameters.
Not sure if I understood it correctly because I donât know the ins and outs of FSRS yet, but to me it sounds like maybe you optimized the preset after returning the âgraduatedâ cards to original decks, and you used âoptimize current presetâ with âpreset:âyour preset for the difficult deckâ -is:suspendedâ, so this preset only has the stats of the remaining âbad cardsâ or underpractised cards to work with. The optimization of this deck wonât necessarily get better the more reviews you do, because its pool of stats isnât growing and itâs always limited to the cards youâre having trouble with. iirc when cards leave the preset, their review history is no longer considered part of this preset. Someone correct me if Iâm wrong.
Yes, the deck has only the leeches, But after the optimization all the progresses I made with them had gone. After ten times with two days in a row it went up to four (and repeated that a number of times). That is slow, but expected with 95% and leechy cards. But after optimization, it dropped down from four to two days again - I hadnât seen longer runs than 10 before, now it is the standard. Maybe indeed it looks at the whole deck and derives from that the behaviour with one single card - I actually had expected that FSRS looks only at the individual card and its behaviour, not the deck around it.
It isnât dependent on the deck but rather the preset; so if Deck A and Deck B both have the same preset, FSRS will consider your performance of both decks and the params for both decks will be modified accordingly.
OK, the preset sets the parameters for FSRS, and it has to depend on them. But does it take its statistics on a deck (or preset) basis or does it do that on every individual card?
Every persons memory is different. But initially, when you start to use FSRS for the first time, it has no idea how you remember and forget things. If you start reviewing your cards eventually there will be quite some data which accumulates.
This data contains things like this:
When was the card due?
When did you review the card?
How did you grade that card (failing or passing; if passing, was it hard, good or easy)?
Using this info, FSRS can generate a model which describes your memory. This model is represented by the FSRS parameters. Now, that model is highly valuable; it describes how your memory functions and is the basis for the scheduling.
But only knowing your memory is useless. Every card is different. One might be very easy, the other very difficult to learn. Your memory overall is the same, of course, which is why the model is the same as well. But every card needs to be considered individually. Which is why, when you review your card, the model of your memory AND the review data of that specific card is used to determine the best interval for that specific card; ensuring you see the card just early enough to remember it, based on a percentage chance (this is dependent on your desired retention).
This is what the algorithm is basically doing; and this is based on sciences (including neuro sciences and psychology). If you truly want to completely separate your leech cards from the other cards (which is potentially decreasing FSRSs ability to create the best model of your memory), then you can use different presets. You can review the link I sent above to read about presets.
Yes, I have one preset for the leechy cards, which I use for a number of decks (I thought it would be a good idea to have different decks for the kind of error that made each card a leech, but those use a common preset,) OK, there might be a risk in separating the bad cards, but I donât want to litter the mainstream with them. (There, btw, I go with only 80%). Probably I need a better understanding on how FSRS works to improve my strategy, avoiding being overwhelmed and at the same time make some progress with the good cards.
What I did now is manually calibrate the desired percentage so it will end up with 2-day-runs of length five. I had to go as low as 83 percent.
What I also did is split the presets for the leeches (which I collect in differend decks by leech reason), because I expect different behaviour for the various reasons.
