This post contains a referral link.
While browsing the Anki forums, I came across an interesting feature request.
It said, “I’d like a feature where AI analyzes my personal learning profile and suggests scheduling settings.” When I read it, my first thought was, “Wait, doesn’t FSRS already do that to some extent?” But when I looked into it, it turned out to be a deeper topic than I expected.
I’ve used Anki every day for 8 years, and I actually use FSRS and run its optimizer, so I took this as a good chance to look into it properly, including FSRS itself.
What exactly is FSRS?
FSRS (Free Spaced Repetition Scheduler) is an open-source algorithm for scheduling reviews that Jarrett Ye of MaiMemo started developing in 2022.
Anki built it in starting with version 23.10, released in November 2023. It was introduced to replace the older algorithm Anki had used until then, called “SM-2.” The latest version is FSRS-6, released at the end of 2025.
The numbers alone might not mean much, but what made me go “I see” when I learned about the behind-the-scenes development was the amount of data used to train it.
Under its privacy policy, the developer of Anki has a way for users to share their review logs for research. From those, a huge set of about 700 million review logs from about 20,000 collections was extracted, and that became FSRS’s training data.
By the way, those “20,000” aren’t the total number of Anki users, just the size of the sample extracted for research. Anki has been a go-to app since 2006, and its actual users number far more. The point is that it isn’t an algorithm built on one person’s gut feeling: it’s built on a huge amount of real data about how people forget.
How it works: three numbers track each card
Looking into how FSRS works, I found that it keeps three values for every single card.
- Difficulty: how hard that card is for you (1–10)
- Stability: how many days it takes for your chance of remembering to drop to the target (90% by default)
- Retrievability: the predicted chance that you can recall it right now
What’s interesting is how Difficulty is handled: as you keep getting a card right, it gradually moves back toward the average. With the old SM-2, a card once judged “hard” could stay stuck as a hard card forever (a problem known as “ease hell”), but apparently this mechanism makes that much less likely.
Retrievability, unlike Difficulty and Stability, which only change at the moment you review, keeps dropping continuously as the days go by. It’s apparently like turning the forgetting curve directly into a formula.
The “Optimize” button was personal optimization itself
After looking into all this, the thing that made me think, “Oh, this does almost the same thing as the forum request,” was the “Optimize” button.
It reads your entire review history (when and how many times you pressed Again, Hard, Good and Easy) and recalculates 21 parameters into values tailored to you. In other words, the exact concern in the forum request, that “even with the same deck and the same algorithm, everyone forgets differently,” was already built in as a feature.

Here’s a quick rundown of what the buttons mean:
- Desired retention: the target for how likely you want to be able to recall a card. Lower it and you’ll review less but forget more; raise it and the opposite happens
- Help Me Decide (experimental): suggests a Desired retention value that fits your study style
- Optimize Current Preset: optimizes only the preset (group of settings) you have selected
- Optimize All Presets: optimizes all the presets you use for different decks at once
- FSRS Simulator (experimental): simulates how your future number of reviews would change if you change the settings
- Reschedule cards on change: a switch for whether to recalculate the next review dates of existing cards when you change the settings
- Check health when optimizing: a switch that also checks the health of your data when optimizing (it takes a little longer when it’s on)
But while researching, I also found something to watch out for. Apparently, if you optimize when you have too few reviews, you can easily get strange results. As a guideline, it seems best to wait until you have at least 500 reviews, and ideally 1,000–3,000 or more. On the forums, I found several reports of settings going haywire right after optimizing, like “the interval for a new card suddenly jumped to 1.4 years” and “the same 2-day interval repeated nearly 10 times.”
As for how often, the general view seems to be that you don’t need to do it every day: about once a month, or each time your number of reviews doubles (100 → 200 → 400…), is enough.
So what’s “new” about the forum request?
After all that, I was tempted to say, “FSRS already solves this.” But when I reread the forum request carefully, I noticed it was talking about a slightly different layer.
FSRS’s Optimize just recalculates 21 numerical parameters behind the scenes, and all you see is “Optimized.” The forum request, on the other hand, asks for AI to translate those numbers into words people can read: explaining and suggesting things like “Here’s your current review load, and here’s how this setting affects it.”
In other words, the personal optimization itself already exists, and what’s being requested is a layer that explains the results to humans. For people who aren’t comfortable with numbers, I think that’s a meaningful request in its own right.

My own take: what I care about more than scheduling
From here on, these are my own thoughts after using Anki for 8 years.
I think having AI suggest settings is a good idea in itself. But if it changed everything at once, I’d worry that it would be hard to go back to how things were. In my case, while writing blog posts, I flag the words I couldn’t use well, review them together later, and rewrite the cards or add pictures if needed. Rather than automating everything in one go, fixing things in small units while checking them seems to suit me better.
But there’s something I care about even more.
No matter how well AI optimizes your review schedule, how often a word is actually used in real conversation is a separate question. You can memorize academic words, or words that only show up in newspapers and the news, as efficiently as you like, but when you actually use them in conversation, people may think, “That sounds a bit unnatural.”
So personally, I think it might actually be more valuable than optimizing the review schedule if AI told me things like “native speakers use this word a lot / not much / it’s on the academic side,” rather than “please use these settings.”
That said, this isn’t simple either. Even in English, the words and expressions people use change from country to country and region to region, so just deciding what counts as “commonly used by native speakers” would be quite a hard problem.
Summary
One thing I gained from this research was learning that FSRS does “personal optimization” far more thoroughly than I’d thought. The forum request is about wanting a layer that translates the calculations behind it for humans, and I can understand that direction.
But honestly, what I’d want more than schedule optimization is information like “How much is this word actually used?” If Anki or a third-party AI tool develops in that direction, I’d really like to try it.
If you haven’t used FSRS’s Optimize yet, I recommend trying it once you’ve built up a fair number of reviews. The feeling of it adjusting to your own memory habits is subtle, but I think it definitely makes a difference.
🃏 I’ve also put all my Anki posts together on one page, sorted by what you need: getting started, settings, keeping it up, automation, audio and troubleshooting.
Referral link: if you sign up through it, I may receive free lesson minutes.
🧠 Anki for remembering, conversation for using
Once FSRS has tuned your review intervals to you, it’s worth making a place to use the words you’ve learned. I try out what I learned in Anki in my Cambly lessons, and “How much is this word actually used?” is something I can ask a native tutor directly. Through my referral link, you get 10 extra minutes when you buy a trial. I’ve been taking lessons on Cambly since July 2021.
Try Cambly with my referral link →
The referral program isn’t available in every region.
☕ Enjoyed this post?
I'm a Japanese blogger who spent 8 years relearning English from scratch.
If this post helped you, a coffee would mean a lot — thank you for reading!





















Leave a Reply