About minimum recommended retention (CMRR)

hello all,

yesterday, thanks to your advice here, I cloned the preset for each of my decks and re-optimized FSRS individually. my true retention rate differed a lot between decks and the idea was to make FSRS more effective for each deck. it brought a lot of changes in parameters. good !

however, I then checked what the minimum recommended retention now was for each deck and the result is the opposite of what I thought. let me give you two examples.

English :
true retention rate - 95.6%
desired retention rate - 90%
minimum recommended retention - 92%

Mandarin :
true retention rate - 88.2%
desired retention rate - 90%
minimum recommended retention - 78%

I was thinking that FSRS was trying to make the true retention rate and the desired retention rate match. on my English deck, I guess right almost too often (my true retention rate is much higher than my desired retention rate) so I think the interval between reviews should grow bigger. having a higher desired retention rate would have the opposite effect, more and more correct answers.

with Chinese, I think the review interval should slightly diminish so that my true retention rate reaches my desired retention rate. if I put my desired retention at 78%, the interval will grow and my true retention rate would dwindle even more with even fewer correct answers

can someone tell me what I didn’t get right ?

thank you.

You are right that the FSRS algorithm is trying to make your “True Retention” match your “Desired Retention”.
It will get better at doing this over time if you optimise your FSRS parameters again as you do more reviews.
N.B. The algorithm is not perfect and the two numbers will probably never match exactly.

Choosing “Desired Retention” is making a trade-off between the amount that you will learn, and the amount of work you have to do.

Higher desired retention = Learn more, but higher workload
Lower desired retention = Learn less, but lower workload

The point of using spaced repetition is to try to optimise the amount you learn per the time you spend learning.

This means that ideally you do not want to increase your Desired Retention too much because there are diminishing returns (significantly increase your workload, to not do much better).

You might want to lower you Desired Retention slightly because you think it is a better trade-off (significantly reduce your workload, to only do a little worse).

However there is a danger that if you lower your Desired Retention too much you end up forgetting so many cards that you are actually learning less per time spent.

This means there is a sweet spot of “good” Desired Retention values that you ideally do not want to go outside of.

The FSRS tutorial has a nice example graph showing this (N.B. this graph would be different for you):

Compute Minimum Recommended Retention (CMRR) is just trying to figure out that bottom bound, which it is not a good idea to set Desired Retention below.
You can continue to use a higher Desired Retention if you want to learn more, but CMRR is telling you that if you want to reduce your workload it is safe to reduce your Desired Retention as long as you do not go below the number it gave you.

See:

2 Likes

thanks I got this.
so for my Chinese deck, does CMRR say that even though I’m not getting a lot of good answers, I should work even less (lower desired retention rate) and it wouldn’t make much difference in terms of results ?
my learning would be more effective and even if my true retention rate will dwindle even more, it won’t be that bad because (that’s probably what I didn’t get in the first place), my graph is VERY different from the average graph you showed.

for English, even more intensive learning (a higher desired retention rate) would be more effective because I would spend a little more time but I would get even more right answers ? for this one, I’m a little skeptical. how could my sweet spot be more than 95.6% ? like 98, 99% ?
and then if my desired retention rate is only 92, my true retention and my desired retention wouldn’t come closer together.

I still don’t really understand and it still looks seriously counterintuitive to me.

CMRR is not necessarily saying you should reduce your desired retention. It is just telling you that it is safe to reduce desired retention to that number.

You can keep desired retention at a higher value if you are fine with the workload and want to remember more cards.

You might want to reduce desired retention (and your workload) for many reasons. Maybe you want to reduce the chance of burn out, or use that time to study in a different way (e.g. listening to something in Mandarin)

I’m not sure what graph you are looking at becuase I don’t think you can generate that kind of graph within Anki.

If you set your desired retention to 92% FSRS would still be trying to get true retention to 92%, not >95.6%, even if it doesn’t manage it.

In my personal decks CMRR has only ever recommended <90% so I’m not sure what is going on here.

Maybe someone who knows more about FSRS like @Expertium could give you a better explanation.

2 Likes

You seem very confused, so I’m not sure where to start.

First of all, desired retention is what you ask for, and true retention is what you get.

Second of all, you have to understand that CMRR is just a recommendation. It just tells you what a good value of desired retention would be. It doesn’t affect anything if you just ignore it. If you click “Compute” and then once it’s done ignore it, nothing will change. Using CMRR is completely optional.

Third of all, CMRR has nothing to do with true retention, at all, whatsoever. You said, “how could my sweet spot be more than 95.6%”, which to me sounds like you think that CMRR is peeking at your true retention. It is not.

Fourth of all, I have no idea what graph you are referring to in “my graph is VERY different from the average graph you showed”. Anki doesn’t plot the workload-retention graph (though this may change in the future).

@rossgb don’t link to that tutorial, just link to the Anki manual. The days when Anki manual didn’t have FSRS-related stuff are long gone.

1 Like

sorry, let me try to clarify: I was talking about my presumed personal graph, what my personal graph for each deck would look like. if for a lot of apprenticeships, a retention rate of 85% to 90% is the best in terms of workload efficiency, for others, it may be very different. and this is what I infer from my CMRR in English and Chinese.

Ah. Yeah, it’s definitely possible that your graph is different. I hope that Anki will add the functionality to plot it in some future release.

1 Like

I’m getting even more confused. doesn’t my true retention rate influence my CMRR ? if not, what does ?

CMRR runs simulations based on your FSRS parameters + a bunch of other stuff, such as “how many seconds does this user take to review one card when he presses Again” or “how many seconds does this user take to review one card when he presses Good”. It runs simulations for different values of desired retention, calculates the “amount of time spent divided by the amount of stuff memorized” ratio, and looks for the value of desired retention that provides the best ratio.

1 Like

oh nooooo, you’re telling me that the answer time is taken into account. I read on one of my previous posts that no, it was not.

I always do other things while working on my decks so sometimes, it takes me 1 minute (or even 30 minutes) to hit “good” when I knew the answer right away.
and it is happening right now. I’m writing this message and my Anki is open and asking me whether I know the word “source of income” in Spanish. shoot !

do you think I should try to stop doing several things at the same time to make Anki’s algorithm and calculations more accurate ?

Answer time is taken into account for simulations, it’s not taken into account for scheduling. When you do your reviews, the answer time doesn’t affect the length of the interval. Answer time is used only by CMRR and the simulator, and they use robust statistics to make sure that outliers don’t screw the results up.

2 Likes

I think it was me who told you that. As expertium explained, it’s not affecting your scheduling at all.

It does affect simulation, but IMO if the relative times for easy/hard, learn/review cards aren’t skewed too much it shouldn’t be a big deal (source: sorry, no source, take it with a lot of salt).

2 Likes

Also, there is a hard limit of how much time gets counted as “review time” per card, which is 60 seconds by default (see deck options > maximum answer seconds).

If you exceed this many seconds (e.g. because of multitasking), then the simulator will just pretend that you used 60 seconds instead.

Of course you might have answered it in less than 60 seconds, but logging 60 seconds is far better than logging 30 Minutes.

2 Likes

exactly, and thank you for the clarification.

I said that because if CMRR is telling me that maybe I should increase the desired retention, it means that my true retention rate is below the sweet spot where the time spent-workload ratio is optimal. do even more reviews, you’ll memorize even better and you’ll be even more effective.

if CMRR is telling me that lowering the desired retention rate would be a good idea, it says that I’m working too much and if I spend less time on my reviews, the time spent would be more effective (far more effective in the case of my Chinese).

can I ask you what CMRR usually says when you’re doing very poorly and when you’re acing it every time ?

in other words, if you can’t memorize stuff at all, should you work even harder to try to make your brain understand that it has to find space to store the information or should you relax because overloading is not a good idea (I can understand both strategies).

if, on the contrary, you’re the perfect information sponge, should you stop reviewing altogether because you’re remembering absolutely everything anyway (CMRR at 0%) or, and that is what I infer from CMRR on my English, should you do even more reviews (CMRR at 99%) because it doesn’t take you a lot of time and you’re getting so many good reviews that the faster you’ll have everything done, the better ?

1 Like

Typically, CMRR is higher for easy material and lower for difficult material.

OK, I thought it was the opposite. I think I’ll create another post because this is super counterintuitive to me.

OK, the previous thread made me understand that as a general rule, the easier the learning material, the higher the CMRR.

but on past versions of Anki, most people were saying that, in most cases, the best desired retention rate was around 85%, maybe up to 90%. I saw people lowering their desired retention rate when their true retention rate was too high (eg above 90%) to try to make it hit 85 or 90.

easier material means higher true retention rate. but if easier material means higher CMRR, it also means that CMRR is advising you to increase your desired retention rate which will obviously also increase your true retention rate, separating it even more from the 85-90.

my true retention rate on my English vocabulary deck is now 95.9%. my desired retention rate is 90 and CMRR is 92. if I bring my desired retention rate to 92, my true retention rate will reach 97 or something. isn’t that much too high ?

1 Like

but on past versions of Anki, most people were saying that, in most cases, the best desired retention rate was around 85%, maybe up to 90%.

CMRR is better than a “everyone, just use the same value” approach, because it takes into account the quirks and habits of each user. 85-90 may not be optimal for some (a lot) of people.

if I bring my desired retention rate to 92, my true retention rate will reach 97 or something

You don’t know that in advance. The only way to find out is to try.

1 Like

let’s try then, thank you again !