Allow 10% as minimum Desired Retention

[JSchoreels]

Hello,

While I don't think going lower than a certain level is necessarily always a good idea, recently a lot of discussions shown that R ranges from 40-70% might in fact be the "best" in terms on knolwedge maintained vs time spent reviewing.

It does actually work OK, some workload ratio for certain low DR seems to display NaN which could be improved later, but for now I'd already like to open the discussion about allowing this or not.

Personally I'm running a 65% DR deck for now and I don't really have much to complain, so I don't really locking the option to users is really worth it, especially when there's some pretty intense warning when they do so.

CleanShot.2025-12-09.at.12.41.41.mp4

[tomer-gom]

+1

[GithubAnon0000]

recently a lot of discussions shown that R ranges from 40-70% might in fact be the "best" in terms on knolwedge maintained vs time spent reviewing.

Can you link to those many discussions? I don' think I saw such a discussion on reddit or the forums (except for a hypothetical survey on reddit by, I believe, @Expertium).

R ranges from 40-70% might in fact be the "best" in terms on knolwedge maintained vs time spent reviewing.

So essentially: in theory there might be a mathematical point where the amount of time spent reviewing is as low as possible, while maintaining the highest amount of knowledge (obviously considering the DR and thus the fact that higher DR would obviously yield higher Retention, at the cost of more time spent per review).

Is that really desirable, though? Why would one be ready to review cards but only remember around 40 % of them (and thus forgetting 60 %)? I don't really see the advantage. If you're going to be happy with forgetting 60 % of information anyways, why not just remove approximately 60 % of your cards, focusing only on things that truly matter to you (and thus having high retention in those cards), while still keeping the review time pretty low?

Edit:

I don't really locking the option to users is really worth it, especially when there's some pretty intense warning when they do so.

That has probably been implemented because there were multiple people thinking the intervalls were too high (and not realizing the impact of DR).

[JSchoreels]

Can you link to those many discussions? I don' think I saw such a discussion on reddit or the forums (except for a hypothetical survey on reddit by, I believe, @Expertium).

Sure, for example this discussion https://discord.com/channels/368267295601983490/1443273721878937650 was the result of this thread : https://forums.ankiweb.net/t/replace-cmrr-with-workload-vs-dr-graph-more/63234/43?page=3

We can of course re-discuss it here, but basically in very simplified summary, I think the main evolution around this topic has been

• In the past, prior of FSRS-6, Compute My Recommended Retention was output 70% for many people. That 70% was a minimum limit the calculator was allowed to give.
• With FSRS-6 and the trainable decay (before it was .2 as a default value if I recall correctly), more and more people asked about this 70% and we started trying to actually see what was the cause of it.
• As you said, the main reason was about the fact that if you want to maximize the sum of knowledge you have (the sum of retrievability of all your cards) is just better achieved by reviewing as little as possible the same stuff and instead growing your knowledge wide. 2 cards at 50% worth more than 1 at 99% Retrievability.
• Even by improving the Help me Decide (the new CMRR) algorithm, it seems for the vast majority of people that have low decay (so cards go to low R more and more slowly, we're speaking years), lower DR might also be beneficial.

Recently, I did also switch to a lower Desired Retention scheduling and while indeed it's not how I imagined I would use Anki after 2 years of it, I start to kinda see the point : If your point is to widen your knowledge over some specific domain like a language, knowing words with a minimum threshold of 50% allow you to probably take an extreme amount of words (Can be 10x, 100x, depending on your params) at the same cost than doing what you do right now.

Of course, there's certain concerns, like for example, if you never did reviews at low R(etrievability), your Calibration Graph might not be good at it. Yet, now that mine has been trained at low R, I can say that yes, FSRS adapt to it well and my actual retention match my current DR (62-65%, I run a fork)

I think you get the main idea and I have no issues discussing even it more, but to come to the reason why I created this PR, is because while there is indeed some things that might need some "expectation management" in terms of what to expect with low Desired Retention, it is a perfectly viable way of using Anki as for example a side tool of your learning more than the engine of it. For people that use Anki as a secondary tool, they might have a very low decay param, which means they will probably never go to 0% Retrievability even after years of not seeing the card because ... they are already seeing it outside from time to time.

So now the question is : If lower DR is acceptable, how low could it be ? In my opinion, deciding a new "golden minimal value of DR" would be just like in the past we decided that "Under 70% it's non sense". In my case right now I do use 62% and I have 0 complain about it, I'm able to review new words at the same speed I encounter them, and the size of my overall passive knowledge is much bigger than what I had in the past, basically going from 5k words to 7k with ... no workload increase.

Another thing I realized by myself is that having 50% DR doesn't mean you'll know everything at 50%. It means you'll see them only when they reach 50% DR. In my case, my 62% DR translate in reality to an average of 87% Retrievability, simply because a lot of cards having very high stability won't drop under 80% any time soon. Knowing this, low DR strategies might even make more sense.

Now, is 10% really a good idea ? Personally I don't see myself going lower than 50%. But who knows if tomorrow or if for some people, a 30% might be acceptable ? And what after that ? I think 10% might be extreme, but it's a safe way to put some value that should not break how Anki work, while still giving to users the right flexibility for how they use the software.

[JSchoreels]

Update : Without going as low as 10%, I've been reviewing cards in the 40-45% range for the past 40 days and one very nice result is that now FSRS has a more broad precision across a broader R range :

It also means my previous decay that was pretty low, ~0.12, is now more like ~0.32, which means "Help me Decide" do not just give me the lowest R possible (Well, it still gives me <10%, but I guess because there's nothing trained there so any optimism there will lead the Memorized/Time ratio to just be higher over there)