FSRS: A modern, efficient spaced repetition algorithm
github.com
github.com
I started out using Anki to learn French vocabulary. I'd make pairs of cards, with English on one side and French on the other. This started out easy, but became utterly brutal and depressing with several hundred cards in my deck. Too many near synonyms.
I eventually took a hint from Katzumoto's Japanese advice, and started making cloze cards. I'd copy and paste an entire paragraph from an ebook or a web page, and hide just one word. These cards were easy, but also effective.
Then I got lazier.
I'd only hide half a word. Or I'd just boldface a word, and mark the card as a "pass" if I could sort of remember that word in context.
And somehow, these cards actually worked better.
Then I got lazier still. If seeing a card made me grown "Oh, not that card", I'd just delete it. If I missed a card 3 times, I configured Anki to permanently suspend it. If I actually needed to know a word, no worries, I'd see it again soon in a more helpful context. And my French vocabulary continued to grow by leaps and bounds.
I don't think that biggest improvements will come from better spaced repetition algorithms. I suspect the biggest wins will come from improved card formats. And it's surprisingly hard to make a card too easy to be useful.
(Source: 35,000+ Anki reps across three languages.)
Mind linking out to Katzumoto’s Anki advice? Is this this one? https://tatsumoto-ren.github.io/blog/setting-up-anki.html
Something that’s become evident, perhaps due to the varied nature of the content (vs. all language) is how much gravity card formatting has. In decks made by others I find myself frequently re-wording the answer sides of the cards which makes a noticeable difference in retention — answers that are succinct and flow nicely stick better.
So in short, I’d concur. Card format is important. If you find a card that feels awkward somehow don’t hesitate to rework it.
Like, yes. I can just spaced repetition the heck out of the notes on the fretboard. Just hammer my head into the notes on a string until I pretty much just give up. Always an idea.
But I've found it more effective to mix contexts up. Some weeks, I just drill notes because I'm lazy. But then I also start tinkering around for a week with some weird scale, like hungarian scales and such. Some other week, I'm tinkering around with arpeggio chords, some other week with cord progressions with power chords, some weeks I'm trying to replicate some sound or feeling of a song.
And over time, the brain is putting things together. Suddenly there are realizations: Hey, this is just a G-Minor scale, which makes sense for a somewhat solemn song. Oh, this is just a G-Major scale but we skip those. Wait, I remeber half the notes of the scale, isn't this just that pattern? Also, if the scale is like this, couldn't we try playing that?
It's just a weird feeling, because the amount of things I think I know expands at a much lower rate of the things I realize I don't know - or don't have the technique for, lol.
But I'm kind of considering using a spaced repetition algorithm to just poke for 5 to 10 minute guitar things to do to get both the muscle memory and the neural connections about them going.
That said, the notes on the fretboard is well worth the effort of rote memorization. I found the technique in this video[1] worked absurdly well for me. Years and years of only knowing the notes half-assed and not very well on the middle three strings got absolutely busted after doing this for a few minutes every time I was waiting on a build or something. I can nearly always find a note now, I'm not quite to instant recall but have substantially more ability in this regard now.
A place where spaced repetition might be worth it is in memorizing all the triads or keys, which is good to have instant recall on. I have these decently in my head but usually have to "convert" from a nearby chord in C major for oddball ones, which is fine while practicing but not what you'd want for performance.
There's no reason you can't do both: Make flashcards that have a few bars of music on the front (tab or notation, whichever you use), and then actually play the music when the card comes up to figure out how the review should be rated.
I spent a long time just playing the different 3nps shapes and never had them really soak in until I started picking small chunks of them out as useful during improvisation and really focusing on them while playing over loops or just noodling.
What would be really awesome, especially if you're improvisation focused, would be some cards that have a short backing section attached to them - so that you can hear some musical context and then work on some small chunk of information (like 2 or 3 strings of a scale). Something like that would probably the best of both worlds and I bet you could learn a lot of material really quickly that way.
Trying to learn from language resources is painfully slow. Trying to watch TV or listens to podcasts is ineffectively difficult.
I'm building Phrasing[1] to tackle just this problem. We've just launched with all of the base functionality[2], including automatic analysis of native material to surface the most important words, phrases, and quotes specific to that material. We then make it dead simple to create flashcards-on-steroids which will automatically be applicable as flip cards, cloze cards, audio cards, and more.
When it comes to SRS, we:
- use our own fork of FSRS coincidentally (that's why I'm here hey )
- do away with scheduled due dates and actually try to keep the user engaged at a specific difficulty (you'll only see hard cards if you're doing well and 'can afford' it, you'll see easier cards if you're not doing well)
- dose out new reviews constantly, as it's better to make forward progress then to bang your head against all those challenging reviews
We just onboarded our second cohort of users, but we're planning to onboard another cohort at the end of this month. If you're interested you can fill out the beta application
Did you ever feel like you were answering correctly, on close cards specifically, not because of an improved understanding but because you were associating the correct answer with the prompt/excerpt? Would appreciate any advice on how to avoid this!
I actually suspect that associating the word with the context helps, as surprising as that might be.
Let's say I grab a few sentences from an interesting article, and I boldface a word. I'll mark the card as "pass" if I at least sort of understand it in context.
On day 1, I honestly might find the word fairly confusing. I can explain what it means, but maybe the grammar is unfamiliar. Ditto for the first few reviews. But then around day 8 or so, the card disappears until day 20 or 30. And the next time I see that card, suddenly the odd bit of grammar is completely natural and obvious. There's maybe some kind of medium-term memory consolidation mechanism occurring? Something happens when I'm not looking at it.
Seeing the word in context somehow allows my brain to grab on. The human brain contains some incredibly powerful language learning machinery. In adults, that machinery still exists. Even if it's a bit rusty. And that machinery seems to work best on semi-comprehensible speech in a natural context. So think of Anki less as a set of facts you must learn, and more as a tool to distill and concentrate natural language so that you can let your brain work on your weak points. (And as soon as possible, start reading lots of books and watching TV! I learned the hardest 10% of my vocabulary using Anki, and much of the rest from context reading books.)
Also, I find it fascinating that LLMs are trained using a very similar process. Either "predict the next word" or "fill in the blank". Now, they need a lot more input than any human does, but the fact that fill-in-the-blanks works so well in both cases is fascinating.
Here are two of the most interesting experiments I tried with Anki:
http://www.randomhacks.net/substudy/ https://blog.beeminder.com/hieroglyphs/
(The only language I ever bothered to push to a high level was French. Spanish and ancient Egyptian were basically experiments to see how quickly I could pick up the basics, using the tricks I learned while working on French. A language that you can use at a professional level is a bit like a pet; it requires ongoing care.)
-Don't do many repetitions of the same card. Eg. if i want to learn 5 new specific words, maybe I use 6-8 sentences that each use 2 new words, so that each new word shows in 2-3 cards. Then I only review each cards 6-8 times in a 2-month period (normally).
-Use long intervals. I rather forget than learn to recognize the thing. But even if I forget, I almost never click "fail". You learn with retrieval effort, even if you just forgotten the thing (that's science)
-Even so, by the 4-5th rep I start recognizing cards, so I be mindful when reviewing, and force myself to really think the answer (occasionally, writing down the answer, for instance). This clashes with my principle of going fast, so there is deff a balance.
-I don't track "streaks", but if I struggle or I have the feeling that something shows up too much, I archive and create more cards.
Does Anki or any of its plugins support creating a difficulty progression when creating cards?
ie. half-word cloze, then full-word, then multiple choice, then q&a, etc.
I’m not sure how this would work in practice. Like would the next difficulty card be added to the deck once the easier card has reached a threshold?
As for multiple choice and q&a---my practice is to make both Cloze and basic cards for, well, most things---Cloze are good for the early steps and for seeing how different pieces fit together, but basic cards are more demanding, and without them I don't quite feel like I've learned a thing.
(FWIW my primary use is physics & the mathematics behind it, not language learning.)
I do agree with the general idea that laziness and going easy on yourself is good though. I give myself quite a lot of slack when grading my answers, applying a "my understanding of this word is close enough to avoid confusion in practice" threshold rather than some impractical ideal of native-level mastery.
I actually had several custom tools that heavily automated card creation—I could grab a sentence from a web page, or bulk import highlighted phrases from an ebook. Then I had a UI which allowed me to easily highlight an interesting word, and either cloze it, or add a Wiktionary definition on the back. Then I had an Anki plugin to bulk import the cards. This could all obviously be combined into a single tool, and occasionally someone tries.
For my most heavily automated experiment, I used a tool similar to subs2srs to import sound, bilingual subtitles, and tiny screen captures from 4 episodes of Avatar: The Last Airbender. That was a fascinating experience, and I'm still earwormed with the dialogue of those episodes a decade later, after only a couple of months of Anki reviews. (See elsewhere in this thread for a link.)
Unfortunately, I'm not convinced that there's a good startup market for language-learning tools. Language learning is normally aspirational, much like a gym membership. And customers don't have any serious plans on how to reach their stated goals. (Again, like a gym membership.) Duolingo isn't terrible, but I suspect—based on lots of Anki experiments—that it should be possible to build much more effective tools than Duolingo. I'm just not convinced that anyone but serious ESL students would pay for them. Too many genuinely good tools in this space have sunken quietly, despite a user-friendly UI and a good landing page.
I guess I am of the few that would pay for a good tool aimed at serious learners, but as I could not find any (in language space almost everything is a Duolingo clone), I am building myself.
So far I think I have proved my main hypothesis: it works really well for me. It's not magic, but it saves me many hours of learning.
Whether it could work for someone else, is an entirely different story... but on the other hand, I have no big ambitions.
It can do sentence card creation, ereader imports, and vocab tracking, among other things.
Unfortunately, it seems like Duolingo is terrible.
Spaced repetition cards seem to blow it away.
I recommend this interview with Duolingo's founder Luis von Ahn: https://www.npr.org/2020/05/22/860884062/recaptcha-and-duoli...
Explicitly, by writing explanations of various grammatical phenomena on flashcards. Implicitly, with usage examples that demonstrate how the grammar works.
Flashcards can teach you pronunciation.
Explicitly, by writing explanations of how to produce various sounds on flashcards. Implicitly, with recordings that demonstrate how things are pronounced. (Ok, paper flashcards can't do that, but we're not talking about those, right?)
If I need an explanation I just paste it into ChatGPT and ask for a breakdown.
At least with SuperMemo, this is actually one of their main tips.
See also: https://supermemo.guru/wiki/20_rules_of_knowledge_formulatio...
The anki default of 8 relapses before suspending a card is much too high. It took me a while to realize, but dealing with problematic cards takes up -way- more energy than learning new cards... it's not energy/time efficient at all.
Do you have any recommended sources on cloze style cards?
[45k reps here for Korean]
Exactly that, specially for language learning. The stated goal of hitting a target probability is pointless IMO.
Language acquisition is not about memorizing words, but progressive familiarization (eg via retrieval practice).
You have to expose yourself to massive amounts of language, create your own cards fast, review them faster, move on even if you "fail", don't do too many repetitions (6-8 should be enough for >95% of cards)
The most important aspect of an SRS algorithm is to be practical: sensible choices, equally spaced out among them, that allow you to decrease / maintain / increase intervals without ever punishing you for failing.
https://www.supermemo.com/en/blog/twenty-rules-of-formulatin...
It will improve a lot your cards
For background, I contributed to Anki long ago, wrote regularly about SRS in the late 2000s and got deeply into them for language learning for a few years. Later, I mostly stopped using them since I found extensive reading both more interesting and a better return on my time.
I’m interested in the topic but am not new to it.
Geeks like us love to get into SRS algorithms and chase after the 1% improvements with a cleverer repetition algorithm or a better input method. The real failure mode is much more mundane: People give up.
I see it all the time. People cram thousands of cards into Anki in their first month learning something. After a month or two they are so overwhelmed with reviews and the reviews become such a chore that they just give up. (Of course, there are people who manage to add thousands of cards every month for years and years. These are the people who would probably do nearly as well with just a pen and paper. )
Two pieces of advice that make all the difference: Make your cards a little bit 'too' easy and be rather selective with your cards.
When you first make a card you're probably actively studying the subject of the card. You're very familiar with it and it's all in your head right now. The temptation is to make a card that is fun and challenging for you right now. Resist! Instead, make a card that feels slightly too easy, so that in a week or two it will be enough to tickle your memory rather than confuse you.
Also, people dump entire books worth of sentences into Anki. I like to use the metaphor of walking through an orchard: take your time, look around and grab one or two of the nicest apples you find. If you try to grab everything you're only going to slow yourself down. Also, if you pick up a rotten apple (uninteresting/too difficult card) it pollutes your entire deck. You start to resist review your flashcards.
I would suspect if this does appear to be the case that there probably is an optimal algorithm to start with but that all methods chosen here decrease in utility if you stay with the same variables (repetition frequency of wrong answers, repetition frequency of correct answers, time delta) for too long, the same way there are beneficial exercise methods for getting specific results when starting out that need to be varied if you want to keep getting results in that area without a decrease in the quality of the results.
Ali Abdaal claimed in one video that memories are strengthened more when something is only just remembered than when it is easily remembered, so as late card prompt as you can get, before actually forgetting. I don’t remember his source for that.
And there’s simply limited time, a thing which repeats wrong answers more but not as much as you need is wasting your time and a thing which doesn’t reduce correct answers as quickly as possible is having you waste time reciting things you know.
So there seems to be room for ways to rank different algorithms.
The other major failure mode is similarly boring: sometimes it's just not useful to blindly memorize things. The exact problem that spaced repetition solves is pretty niche, and while it can be "close enough" to be beneficial in adjacent endeavors, it's also extremely subject to Goodharting. The more hardcore you go into optimizing your SRS procedure, the likelier it gets that the metrics you're optimizing for depart from your actual goal.
Specially for language: https://thehardway.guide/srs (own content)
For example, learning from books or mathematics?
What even is a good progression for creating cards? Only add them slowly and optimise for having a few cards in your review queue and only add after several days or weeks of no cards? Find a good text and just add them?
Sorry, it appears that the people in this thread are clearly deriving value from SRS and speaking as someone who's tried SRS and Anki on multiple occasions and failed to keep it up, it would be helpful to get insights =)...
At some point I even added some more advanced proofs, which actually go the opposite route of what's recommended here - they were cards which weren't immediately answerable, but required some thought. I experimented with some different ways of doing this - I created my own "Math Proof" notes, which allowed you to write a bunch of steps, and you'd then get one prompt to remember a "proof statement" (what you're trying to prove), one prompt for the proof overview/steps, and then one card for each specific step.
At some point life overwhelmed me, and I couldn't continue with Anki (work + kids). But I've actually been getting back into it in the last month, and having all those cards has been an incredibly great way to relearn a bunch of math I'd semi-forgotten, fairly easily.
I also use Anki for:
- Memorizing Geography stuff (all the countries and capitals).
- Memorizing songs.
- Memorizing random misc facts that I run across (usually things like history, politics, etc).
Useful approach of going through a textbook as well thanks! I'll try it.
After writing out the original message I thought about using ChatGPT to suggest solutions, the execution leaves something to be desired, but the kinds of solutions it's suggesting are interesting.
For example giving it a block of text and asking it to generate questions, use that to find questions I find interesting and then answer them and turn that question / answer pair into a card.
It does seem that figuring how to not fall off the bandwagon is the biggest thing to work out first.
I'm trying to find the best way to do this in 2024. What I used to do is use a plugin called "Cloze Overlapper", which helps generate these automatically - you give it all the lines, and it creates cards where you get one line before, and need to guess the next line. Unfortunately this plugin is behind a paywall or something, unclear if it still works.
There are alternatives today, a popular one is something like Lyrics Cloze Generator, which does something similar but ends up working a bit worse IMO.
Btw, I use this both to memorize songs I randomly feel like learning, but also to memorize songs in another language, as a way of learning more vocab in a fun way. Not sure if it's optimal but I figure I might as well, if it's a song I want to learn anyway.
I think I should have another crack at it starting small with some of these and see where it takes me =)...
I went back and dug up some of my more valuable cards from my last attempt, and some of them were so complicated that I had to break them up into 10–15 cards now that I know how to design good cards.
That said, I don't think the point is to make them easy. Since you want to aim for a 90-something percent success rate to get the most benefit out of it, too easy will make the going slower. The point is to make them atomic and that usually feels easy, even if you fail 10 % of recall tests.
If it looks at all interesting to you [1], I'd love to chat more and hear more about your experience / approach / etc. Feel free to reach out at ben [at] phrasing [dot] app
I currently have a (foolish?) project. I’m trying to memorize the 750 cards of a quiz game that we play in the family. All questions are answered by a year. So for example: “What year did Coca Cola Light come out?”.
I have been using Midjourney to generate images for those cards which makes it so much eaaier to recall.
I have a system. I have a person representing each century; Einstein is 1900-2000 and for example Mari-Antoinette is 1700-1800. Then items represent the decade; a sixties car represent the sixties, jacket with shoulder pads the eighties and so on.
I do something similar for the last digit.
Then I have Midjourney generate such pictures, in a comic book style, and I save the most fun or absurd one and use that on the back side of the Anki card.
The image is often easier to recall than the year by itself
I'm currently doing the same for geography ... flags, capitals, etc. I reckon it will give me a +5 advantage on most trivia nights
1. Look up PAO - it's a tangentially related mnemonic system that uses images.
2. Test your recall periodically with/without the images otherwise you may find that the true nature of recall (aka in the wild) without said images is adversely affected.
3. Is it Trivial Pursuit? A friend of mine deliberately memorized all the cards of the original Genus edition.
The image is on the back of the card, with the answer. So trying to recall the year, I don’t get to see the picture. But, the picture is the first thing that comes to mind when I get the question and the from it I get the year.
Just a few days ago, I published a Python implementation of the classic SM-2 algorithm that I use for https://python.cards, but I may switch to FSRS. https://github.com/vlopezferrando/simple-spaced-repetition
Any chance you are open sourcing the web app? I can see it being a popular way for niches to show flashcards eg. for students to learn X
I believe the hardest part of using spaced repetition to learn programming is creating good decks, it's a ton of work.
About open sourcing the web app, I might do it. It's a Django app, and I've published some videos while coding it (https://www.youtube.com/channel/UCyWUj9r0soytotuuh2JnPrw), so it's no secret.
> I believe the hardest part of using spaced repetition to learn programming is creating good decks, it's a ton of work.
Agreed. To have always been disappointed with the programming decks I have created.
> About open sourcing the web app, I might do it. It's a Django app, and I've published some videos while coding it
I saw the videos, thanks. I think a generic platform would have value, even if there was a paid option with more features (payments etc.)
Anyway, I can sign you up manually if you wish.
Anyways, thanks for the project, it looks nice - I'll be happy to grab the decks when they're ready :)
I think Anki is a great app, I just wish there was an algorithm that would just randomly sample cards (with probability proportional to how urgently you need to review it) rather than put a review date on them. Something like https://github.com/fasiha/ebisu but available as an Anki plugin (if that supports custom algorithms on mobile yet?) or a similar app with an open format for cards.
Because I'm not a student. Why is it so hard to accept that I know what I want and am just looking for recommendations that do that? If they don't exist, so be it. The point of my question was to find out what exists, not to get a lecture on what I should want.
That's how SRS already works. But if you clear out the "backlog" you can no longer practice whenever you want and must wait for a period before you will be able to practice again.
We don't use due dates really (except for some analytics by request), but instead try to keep a user at various difficulty levels.
We care more about forward progress then making sure you're always 100% caught up (if a words really that important, you can either favorite it or you'll see it again).
It's not an open flashcard app, we're really also heavily focused on automatic sentence mining for foreign languages, but I'd be happy to answer any questions about my experience adapting FSRS.
You can see more about the product at https://blog.phrasing.app/phrasing-first-look/
Because of the extensive benchmarking most people probably will not benefit from refitting the weights to their collection until they have thousands of reviews (author recommends 1k+).
Note it still works fine even if you do your cards late, since the recall probabilities are based on the stability and when you last reviewed the card, and the stability will update a bit longer if you somehow managed to still recall a card after the due date.
[0] https://dl.acm.org/doi/10.1145/3534678.3539081?cid=996605471...
[1] https://github.com/open-spaced-repetition/fsrs4anki/wiki/The...
For language vocab, I use Clozemaster, but then I put the sentence into Supermemo after I get it right because the SM algo is that much better. They also have ChatGPT explanations for each word of the sentence. I also put those into Supermeno.
The hardest part is making good cards for sure. This will help with more ideas: http://arxiv.org/abs/2401.01257 and https://rust-book.cs.brown.edu
Also, I have not found a flashcard program that lets you make cards as fast as Supermemo. With Supermemo, you paste in a chunk of text (ctrl-n), highlight a word you want to make cloze deletion from, and press alt-x. You can do it multiple times on the same piece of text, and every time you do it, you get a new cloze deletion card in your reviews. Almost every other app, you have to make a single card at a time. I think its because most other apps stick too hard to the cards and decks metaphor. Supermemo uses a tree structure to organize everything. It makes a big difference. Technically, you don't really need to organize anything though. People act like the material they learn is going to be categorized into neat decks in their head.
> With Supermemo, you paste in a chunk of text (ctrl-n), highlight a word you want to make cloze deletion from, and press alt-x. You can do it multiple times on the same piece of text, and every time you do it, you get a new cloze deletion card in your reviews.
In Mochi it's mostly the same. Press n (new card), ctrl-v (paste text), select word, ctrl-l (make cloze), ctrl-[n] (where n is 0-9) to create a cloze group. Each cloze group will have its own "card".
Aren't there three separate items?
1) Your cards (or a subset of cards).
2) The history of your interactions with them.
3) An algorithm (potentially just-in-time) taking that history into account to present cards to you, thereby adding to the history.
In practice, that's illustrated by the fact that there's now a button to fully recompute all intervals and difficulties based only on your history and the current algorithm tunings. And if you've already been using FSRS and the tunings haven't changed, the recomputation won't have any effect because it's equivalent to the incremental computations after each review.
So in principle it could be thought of as a just-in-time pure function, which involves a cache of generated data only for legacy & performance reasons.
I would love some feedback.
BTW, the code is on github [2].
I'm currently learning MRI technology for my radiology board exam with the following youtube playlist: https://www.youtube.com/playlist?list=PLPcImQzEnTpz-5TzxyyoY...
This is what got me the idea.
So, SRS algo is just part of the solution, and definitely not a silver bullet. There are a lot of things about memory, such as how emotions affect encoding - you don't need SRS to remember your first kiss or graduation day. People learn their mother tongue to fluency w/o using any kind of SRS, while most people fail to learn a foreign language to fluency even if they use SRS.
First and Second Language Acquisition is just apples to bacon. The baby and adult brains are different, and their needs, capabilities, and context are vastly different.
I guess this at least allows me to remove my old anti-ease-hell addon.
> During term time if you set the retention to like 70 % so you don’t spend too much time on it daily and then increase it to ≥ 90 % when the holidays start.
They provide actual cards, you can use the old way.
Or you can then take a picture of them using their free app Scribzee, and review them digitally too. My son just aced his anatomy exam with them.
All are MIT licensed I believe, Anki is primarily AGPL
* https://github.com/ankidroid/Anki-Android/releases/ (Parallel.A)
After following the instructions [1] (joining the Play Store testing programme, then the Google Group), the Play Store should distribute + automatically update the alpha.
Also, was it always free? When I looked at what was available for Anki, the remote solutions were paid only and not selfhostable.
Maybe I didn't look well enough back then.
It would also be interesting to see how SM-18 compares to SM-17 ;)
By the way ,SuperMemo CEO said that they will release the algorithm API. If it’s not very expensive, I will compare it with FSRS.
You can literally ankify (and therefore memorize) anything you want to remember: names of friends' children, algorithms, arguments from papers or articles, code, ...
Question: short code
Answer: output