My hypothesis would be that if you let people use FSRS-5 for a while scheduling the relearning steps and test that data, the errors won’t look so bad.
The FSRS model is based on heuristics. Academics is behind when it comes to optimized models for spaced repetition. You will find 0 peer reviewed sources that can compare to FSRS in efficiency when it comes to scheduling long term reviews.
You’re also seriously just arguing over semantics at this point. In these conversations, short-term memory is not defined as 20 seconds. The term is used to refer to <1 day intervals, which FSRS was not made to deal with.
This is like complaining about someone using the exclamation mark to mean “factorial” in mathematics, because in C programming, you were taught that it acts as the NOT operator. You’re complaining about language here.
Short term memory as defined in this forum has got NOTHING to do with the academic definition, and nobody here cares about that definition because people use Anki in different ways that change how long it is retained within their memory, such as using mnemonics to aid memorization, example sentences, rereading, etc.
3 Likes
Apt name
FSRS isn’t operating in a vacuum separate from academic research. The forgetting curve is based on studies of memory decay, so dismissing academic insights as irrelevant doesn’t really hold up when they’re part of the algorithm’s foundation.
If you think your analogy reflects what I’m saying, then you just didn’t follow my argument. I don’t mind using the terms short-term and long-term in the colloquial way they’re used in this forum. But when people start saying things like “FSRS isn’t designed to model short-term memory,” we need to get specific about what they mean. Are they saying it can’t handle times less than 24 hours because of the integer-type limitation, or are they suggesting that the algorithm fundamentally cannot model it because the forgetting curve ceases to be applicable due to short-term memory functioning differently in the brain? That’s why I’m referring to memory in the academic sense. As far as I know, the concept of long-term memory should cover all the relearning intervals we’re discussing, and the forgetting curve should model these just fine. If that’s not the case, why? I haven’t seen an explanation yet.
Is this based on using the current system where all times less than 1 day are counted as 0 days? I know it’s a nightmare to fix/change this in Anki, but I don’t know if you found some work around just for testing purposes.
This to me implies the answer is the technical limitations and not theoretical limitations of the algorithm itself. Am I getting that right?
My question really boils down to, if Anki used minutes instead of days for its base unit of measure, would FSRS work, or is there still some fundamental issue the algorithm has with shorter intervals?
To be fair, nothing exists in a vacuum.
I don’t think the point is to dismiss academics as irrelevant, only that the constraints of the term as used in that context don’t (necessarily) overlap with the term as used in the context of discussions around FSRS modeling.
It’s definitely reasonable to define a boundary for what is meant by ‘short-term’ (or, really, any term) and, as noted, that seems to be same-day intervals which, as expressed by LM Sherlock, hew to a combination of difficulty in modeling shorter-term forgetting, on its own merits, as well as architectural limitations imposed on the part of Anki as it currently stands.
Someone may correct me, but the limitations seem to be two-fold in that regard. It is not strictly a limitation based on Anki’s handling of intra-day intervals.
1 Like
I agree with much of what’s been said, but I think the focus on short-term intervals (<1 day) often misses the practicalities of how most people use Anki and FSRS. In my experience, intervals shorter than a day become largely irrelevant in practice because I don’t tend to attend the cards the exact moment they’re due. Instead, reviews are handled when it’s convenient, which means the exact timing of those short intervals rarely aligns with real-world behavior.
I think it’s reasonable to view FSRS as a tool for scheduling long-term retention, while short-term learning steps remain a user-managed process, possibly influenced by intuition or a more general approach. This division plays to the strengths of both the user and the system.
Ultimately, I think that discussions about short-term vs. long-term intervals should focus on how the algorithm can meet user needs practically, rather than being bogged down in theoretical distinctions that may not impact real-world outcomes significantly.
9 Likes
The whole point is to optimize short-term intervals according to a short-term memory model and remove the issue of figuring out which learning steps work best, which has long been an issue, mind you it is not only about optimizing retention but making the learning process less strenuous. I suggest you stop projecting your opinion onto others and what other people are doing with their Anki. Anki users are not a monolith.
FSRS has got little left to improve on and this is one area that is very niche. An area that you are probably not even using the feature of in this current build so it won’t concern you or others sharing your opinion.
Thank you for your response, and I apologize if my previous comment came across as dismissive or overly projecting my perspective. That wasn’t my intention, I genuinely value the insights shared here.
To clarify, I’ve actually experimented with leaving the learning and relearning steps blank to allow FSRS to handle those intervals entirely. However, I noticed that most of the recommended intervals ended up being around 1d, which didn’t feel quite right for me. That’s why I turned to the forums and discovered this feature of the FSRS helper add-on mentioned in this post. Interestingly, the learning and relearning steps it recommended aligned quite closely with the steps I was using before leaving them blank.
I absolutely agree that every user is different, and their workflows and needs can vary widely. While I recognize how amazing it would be for FSRS to seamlessly handle learning steps, it seems excessive to fixate on this topic, as the gains from fine-tuning these short-term intervals are likely to be marginal, especially since FSRS already handles intervals of ≥1 day exceptionally well. This is largely because, as I mentioned earlier, I doubt that most users are waiting at their devices to review their cards exactly when an interval passes. Reviews tend to happen when convenient.
That’s why I believe the default presets, which allow learning and relearning cards to be reviewed within a single day, are appropriate (then, FSRS can run its course for longer intervals). If the user feels compelled to adjust those intervals, so be it, and tools like the one in this post can help guide that process. Anki already warns against choosing learning intervals of 1d or more with FSRS. Perhaps additional details in the manual or on official resources could further clarify how to approach learning intervals for those who find it daunting to determine their ideal settings.
3 Likes
There is a workaround. I have enabled it. All that you need to do is put in some custom learning steps. If you have the FSRS Anki helper add-on, you can find out what the recommended steps are, and you can put them into your learning steps. The auto-scheduler would take responsibility afterward.
Perhaps I missed the detail, but out of curiosity what does it indicate if there simply are no steps recommended:
It means you don’t need (re)learning steps.
1 Like
I think in that case, it should tell you that explicitly, with a text like “You don’t need relearning steps.”
OK. I will update the add-on.
Question: How do you select the DR? I think there should be a way to calculate the DR at which average stability increases the most.
1 Like
The minimum recommended retention is your need for this case.
1 Like
Well, given that STM (or whatever we name it) behaves very differently I doubt this would be of much use. Wdyt?
Then I have no idea. If we don’t have a model for short-term memory, it’s impossible to calculate the DR to maximize the stability.
1 Like