Background
With the debut of FSRS-5 in Anki 24.11, there's now considerable controversy surrounding whether FSRS should control short-term intervals. Additionally, some inaccurate information about short-term memory is spreading.
Therefore, I feel it necessary to provide some clarification.
Fact
- In Anki 24.11, when FSRS is enabled and (re)learning steps are left blank, FSRS can control the (re)learning steps when it deems necessary (when the next interval < 12h).
- FSRS-5 was not initially designed to model short-term memory. Its primary focus was on considering the impact of short-term reviews on long-term memory.
- During the optimization of FSRS-5 parameters, short-term review results were not used as labels in supervised learning. Using a next token prediction analogy, short-term reviews appeared only in the input/context tokens, not in the next tokens.
- Benchmarks show that considering short-term reviews improves long-term memory prediction accuracy. However, this doesn't necessarily mean FSRS-5 can accurately predict short-term memory.
- Recent experiments involving short-term review results as optimization labels led to a significant increase in FSRS prediction errors and overly conservative long-term memory predictions. This suggests that long-term and short-term memory patterns may differ, and using a single model to predict both may not be ideal.
- Short-term reviews have a significant impact on short-term memory. But it’s too complicate to model.
FAQs
Most of my answers are based on my open-source research: open-spaced-repetition/short-term-memory-research
What inspired the module considering same-day reviews in FSRS-5?
The inspiration came from my research on short-term review data:
In this graph,
r_history represents the history of review ratings, where 1 indicates 'again' and 3 indicates 'good'.Clearly, in short-term reviews, more 'again' responses lead to lower long-term memory stability.
Conversely, more 'good' responses result in higher long-term memory stability.
Therefore, in FSRS-5, if you rate a card as 'again' during short-term reviews, the memory stability will decrease. On the other hand, if you rate it as 'good', the memory stability will increase.
How did you conclude that short-term reviews significantly impact short-term memory?
This conclusion is also derived from my short-term memory research data:
In short-term reviews, memory stability gradually increases: 1.87 minutes → 13.88 minutes → 6.26 hours → 1.08 days
The growth factor here far exceeds the default ease factor of 2.5 in SM-2, which leads me to conclude that short-term reviews have a significant impact on short-term memory.
Why allow FSRS-5 to intervene when users leave learning steps blank?
This issue has a complex historical background. For details, please read this discussion: Graduate new card when the user presses again or hard and has 0 learning steps - Anki / Scheduling - Anki Forums
Initially, I observed that when learning steps were left blank, Anki still added a default step, which differed from the behavior of blank relearning steps. I believed this was incorrect; a blank learning step should logically skip short-term review and proceed directly to long-term review.
However, this had a side effect:
if the initial stability ofagain,hardandgoodis shorter than 1 day and the desired retention is 90%, the intervals of those three buttons will be the same.
Someone suggested:
I may be off base here, but I’m assuming what people really want is for FSRS to do the scheduling as optimally as possible without any inflexible learning steps getting in the way. If so, then when the stability is less than 1 day, could we not leave the card in learning and schedule it exactly according to the stability?
This led to the Pull Request: Let FSRS control short term schedule by L-M-Sherlock · Pull Request #3375 · ankitects/anki
Throughout this process, I never suggested that anyone should leave learning steps blank. I was simply trying to optimize the experience for cases where learning steps were already blank.
How should I set learning steps then?
I recommend referring to the recommended settings in the Steps Stats of FSRS Helper. These settings are based on your Anki statistics, not on any short-term memory model (except for the forgetting curve).
However, please note that by design, it can recommend at most two learning steps and one relearning step. Also, due to some limitations in Anki's learning steps, it cannot fully meet the desired retention. For more details, please see FSRS Helper - Recommended Steps - Anki / Add-ons - Anki Forums
If FSRS Helper can recommend learning steps, why not integrate this into the FSRS model?
FSRS Helper's Steps Stats are not based on any short-term algorithmic model. This means it lacks generalization ability (for example, it can't recommend a third learning step based on the first two recommended steps), let alone integrate with FSRS's long-term memory model.
Additionally, what I didn't mention earlier is that FSRS-5 can't detect your adjustments to learning steps. It will only adapt in the next optimization after you've accumulated more review data under the new learning steps. Therefore, I also don't recommend making significant changes to your learning steps.
What is your current progress in short-term memory model research?
Unfortunately, there's been little progress. The spacing effect, which is very important for long-term memory, also shows up in short-term memory, but its effect doesn't always grow steadily with time. Also, short-term memory data sometimes goes against the forgetting curve: retention rates can increase over time instead of decreasing.
If you're interested in this research, please check out my repository: open-spaced-repetition/short-term-memory-research
Key Takeaways
- FSRS-5 primarily models long-term memory but considers the impact of short-term reviews on long-term retention.
- Short-term reviews significantly affect short-term memory, but modeling this is complex and a comprehensive short-term memory model is not yet available.
- In Anki, if you previously had non-blank learning steps, it's not recommended to switch to blank steps when using FSRS. Maintaining appropriate learning steps is still important.
- FSRS Helper can recommend learning step settings based on personal statistics, offering a data-driven optimization approach.
