Does the ease lasts for years

Eric's Avatar


28 Mar, 2017 10:15 AM

I was wandering if after finding the corrdct ease, it lasts so for years, and you keep hitting GOOD, or does it change over time?

  1. 1 Posted by Vit on 28 Mar, 2017 05:56 PM

    Vit's Avatar

    I do not want to speculate ...
    What is the real question ?

  2. 2 Posted by Eric on 29 Mar, 2017 06:35 AM

    Eric's Avatar

    Thanks for replying.

    What stands in the background of my question is that I don’t want to use any of the existing SRS, because they are all based on Wozniak's way to know if the EASE can be increased, which is to aim for non-perfect retention. That’s good for normal learning, where the knowledge can be accessed upon questioning. But what I'm learning has to come to mind without questioning, I need to associate bits of knowledge from all over a huge vague and complex corpus. I can't even reformulate the knowledge to Q/A format, because I don’t fully understand it, and because I don’t know which part is significant. I want perfect retention of that stuff so I can unscramble it.

    So I want to write myself a little Excel macro instead, which will use only Boolean feedback ("I remember/I forgot"). But how can my SR algorithm tell about the EASE? If the EASE doesn’t decrease over time, it can increase the EASE after every positive feedback, until it reaches a negative one, and then forever use the previous EASE.

    And this is how I got to my question, what happens to the EASE in the course of time?

    I would greatly appreciate you experience on that matter.

  3. 3 Posted by thomas.tempe on 08 Oct, 2017 01:47 PM

    thomas.tempe's Avatar

    Hi Eric,
    Ease stays the same as long as you only hit "good". It changes only if you hit the other 3 buttons. It starts high, and gets decreased as you fail cards. If you want more spontaneous knowledge of a given card that you remember just so-so, just fail it once or twice, and it will do. (I'm doing just that).

    However, your use case seems quite unorthodox. For starters, remembering things you haven't understood is terribly difficult. If the corpus is such that you need to understand some bits before you can fully grasp other bits, then incremental reading might be what you're looking for.

  4. 4 Posted by Eric on 24 Oct, 2017 01:50 PM

    Eric's Avatar

    Hi Thomas, thanks for replying.
    I actually meant another question: I want to know if the algorithm really works. Does the human memory really behave like that, that every item has a fixed EASE, and the software manages to find that real EASE?
    If so, a user should find himself remembering things quite well, and keep hitting GOOD for years.
    Is it really so?

  5. 5 Posted by thomas.tempe on 29 Oct, 2017 02:25 PM

    thomas.tempe's Avatar

    Hi Eric,

    that's actually an interesting question. I've been using
    Anki for over 6 years now, and here's my understanding:

      * First of
    all, I've got close to 14 000 active cards in my deck, mostly difficult
    ones, and a Correct rate of close to 90% (benchmark is 90~95%). If that
    was your question, then yes, SRS is definitely working.

    Then, on a more
    detailed level:

      * your recall performance is actually quite variable
    from time to time. I got into a habit of, when I can't think of the
    answer for long-term cards, to pass them on the 1st try (rather than
    fail them). Quite often, they're obvious the next day; a bit like when
    you can't remember a word in your mother tongue.
      * your criteria for a
    "remember" are also quite variable across the years. At first, you might
    expect instant&obvious recall, and fail anything below that. After a few
    years, the topic might have fallen off your radar. And what to say of
    clozes where your hid a single word, but what you really wanted to
    remember was a concept?
      * you should definitely not be happy with bare
    SRS. I'm trying my best to build mnemonics for each item I learn. (I
    learn mostly Chinese vocabulary, but also trivia, people's names, and
    reading notes). For Chinese, I don't let myself learn a word if I
    haven't "built a relationship" with that word. Eg: looked it up in
    Google Image, or Wikipedia, or imagined a context where I'd want to use
    it. After a while, the mnemonic fades into oblivion, but the word
    becomes active vocabulary. Sometimes (rarely), I find myself having
    forgotten the mnemonic, and the card as well.
      * Your own ease of
    remembering is not set in stone either. Eg: the first 50 Chinese words
    might be quite a pain to force into your brain, but once you're hitting
    1000 and have a firmer grasp of how the language works, all Chinese
    words (including the first ones) have become much more obvious. If you
    hit a leech, let it sleep for a year, and it might have become easy,
    just because of the other things you've learned around it.
      * Most
    importantly, if it's something you use in your life (like vocabulary, or
    work-related factoids), after a few weeks or months, it gets to the
    point where you recall them more frequently from your normal activities
    than from SRS. At that point, SRS stops bringing value (but it's not
    much of a burden either). I've had the case of failing a bunch of
    work-related cards a year or two after quitting that job. Also, I'm
    currently living in China, and I fear that when I move away, a lot of
    the vocabulary I've built will suddenly appear to be much harder than
    Anki thought it was.

    All in all, we're looking at a very noisy process.
    The ease factor is most likely artificially low for a lot of the cards
    in any large deck. And that's fine, too, because by the time a card has
    been in your deck for a few months, the cost of keeping it there for a
    few more years is very low -- much lower than the cost of re-evaluating
    its actual ease.

    So a more nuanced answer to "does SRS work?" might
    actually depend on what you're trying to learn, and what timeframe
    you're in. For the few long-term factoids I have that I don't use much
    in my everyday life (geography, reading notes, obscure vocabulary...),
    my subjective assessment is that it works quite well; but I haven't run
    the math. For the bulk of my deck (Chinese vocabulary), my goal was to
    get to where I could use it in real-life (say, 1~6 months); for these,
    Anki was a key enabler; definitely working.

    Have fun,


    2017-10-24 21:50, Eric wrote:

  6. 6 Posted by Eric on 07 Nov, 2017 07:12 AM

    Eric's Avatar

    Thank you Thomas, these are very interesting observations.
    To sum up your answer, if I get it right, the grades can't teach us about the existence of a fixed "ease factor", because even if there is one, the grades are also affected by blackouts (observation 1), subjective grading (obs. 2), knowledge depth (obs. 3), familiarity with the subject (obs. 4) and unreported revisions (obs. 5), which are all inconsistent (and there are probably some other factors we haven’t discussed). And therefore, regarding my question, using the biggest "ease factor" that never failed might inflict too much revisions.
    But besides answering my question, your observations tell us something deeper. Every predicting algorithm has to be based on some model, and these observations show that Anki's algorithm (as well as all the SM algorithms and their derivatives) isn't based on a mathematical memory model, because it doesn’t take into account all the factors which might affect the grades. The algorithm is apparently just common sense that you should review periodically to refresh your memory, and that the more you review the longer you remember, and the "ease factor" is merely the periods' growth rate. And since a fixed growth rate was introduced by Pimsleur in the 60', and increasing the interval upon a test was introduced by Leitner in the 70', I'd say that SM's innovation (introduced in the 80') is to adjust the growth rate according to the tests' grades, even though the grades are probably also affected by other factors. So advertising SM as a "scientific conquest" is a little windy.
    Actually, I'd be fairly surprised if there really is a fixed growth rate, because the number of repetitions increases in a decreasing rate (it increases one at time, so the ratio is 1, 2 (X2), 3 (X1.5), 4 (X1.333) etc.), it's reasonable that every consecutive repetition is less influential. Come to think of it, Anki itself uses a decreasing growth rate in the default settings, since the first intervals are 10 minutes, a day (X144) and 2.5 days (X2.5). But on the other hand, 4 out of 5 of the factors you mentioned should increase the growth rate, so it's hard to guess what's really going on with the intervals inside our brain.
    But the bottom line is, you say that SRS is definitely working, even if the grades might fluctuate a bit and never settle on the "right" ease factor, so as to achieve the "good" button on every test. I'll strengthen your argument a little. You wrote in the first observation that you got into a habit of "passing" long-term cards (I guess you meant the "bury" option) and remembering them the next day. Why not short-term cards? Also, after no one answered my inquiry here, I went on to analyze the Mnemosyne-project's database, and from what I figured, people usually really did forget their cards. Have in mind that the Mnemosyne DB is young and I found almost only short-term cards. This difference between young and mature cards might imply that after a while the algorithm really does find an effective repetition frequency, so a young card, which hasn't find its rhythm yet, usually fails due to late repetitions, but a mature card, which has already gotten to the right pace, usually fails due to blackouts.
    So in conclusion, it looks like a Boolean feedback algorithm isn’t such a bright idea, because it doesn’t let you fine-adjust the growth rate, so I should stick with graded feedback. But hopefully I'll refer to that in more detail on

    By the way, your previous comment on incremental reading is a good advice, and that was my plan all along. But not the SM way of chopping it down to atoms gradually, rather just read some text in increasing spaces according to how well you recognize it, as Damien advised here:

Comments are closed, but you can start a new discussion.

Keyboard shortcuts


? Show this help
ESC Blurs the current field

Comment Form

r Focus the comment reply box
^ + ↩ Submit the comment

You can use Command ⌘ instead of Control ^ on Mac