Cheating the SRS

Yes, yes, I know. There’s a Search function. I’ve used it.

Now, on to the questions:
Following the @MegaZeroX thread pointing out that there is no such thing as cheating the SRS, and essentially claiming that reviewing items before the lag period is over is not detrimental to learning, would it be fair to say that perhaps growing intervals are not useful?

After all, if it is the case that seeing an item after 2 months when the lag is meant to be 4 months is not detrimental to learning, why not simply have a fixed lag of 2 months?

And further, why even 2 months? Why not 9 and a half weeks and 21 hours?

And how would we even determine the optimal lag(s) per individual? Seems like heuristics with AI is the only efficient solution I can think of.
But then again, if we were looking for vague non-answers like that, perhaps we could just pack up and go drink tea instead.

From experience, I don’t think a 2 month lag right off the bat would work. After all those first repetitions seem to have an effect.
But at what point should the lags stop growing then?

Note: I know how these discussions can devolve rapidly so a couple of points to try and keep this train on its tracks may be in order.
Am I saying that SRS does not work? No.
Am I saying that growing intervals don’t work? No.
Am I saying that Donatello is way better than Raphaël? No. But that would be true regardless.
What am I claiming? Absolutely nothing. Just asking questions.



I think those interval are commonly accepted as part of research by Ebbinghaus on Human memory and his establishment on the “forgetting curve”.

How to take advantage of the forgetting curve - Quora

Basically if you review before, it will be considered as refreshing your memory (=1 Repetition). There’s no harm, on the contrary. It’s just that they are trying to optimize the retention VS time spend per new item learned. That’s why they put those interval. But if you want to review before, sure that can only be beneficial ! :slight_smile:

Also you can see from this curve, the percentage of information retained tends to flatten over time as repetition increases. Making knowledge stick as long term memory as opposed to short term memory. It wouldn’t work if you review 5 times in 2 days, versus 5 times in 3 month.

Disclaimer: I haven’t read Ebbinghaus original work, so I don’t know how much was theorized by him and how much was later completed by others.


From the few things I read, I feel like Ebbinghaus used to have an empirical approach. So probably the number were based on the observation of people and probably some statistic. However I have no idea what were his hypothesis concerning sample size and limits. That’s to say, there might well be a variation from an individual to another, but by what margin, I cannot answer

1 Like

I think that growing intervals are the most important part of something being an “SRS” and not just any old flashcard app. The idea is that we want to spend as little time as we possibly can in flashcard review, so we want the gap between each time we see the card to be as long as we can make it without it being so long we’ve forgotten the information. All the empirical studies and algorithm tweaking are just trying to figure out the most efficient set of gaps.

It’s possible that if our brains were wired differently then “fixed gap” would be necessary for retention. Luckily we can get away with longer gaps, or we’d end up spending all our time reviewing flashcards…

1 Like

Ah yes, the forgetting curve. So here’s a question pertaining to that: has anyone seen anything about what happens beyond the 5th repetition? I’m curious to see how large the lag can be before the retention just plummets. That is to say, what if after the 3 month lag we do the next one at 9 months? What if instead we do it at 3 years? At what point is the lag simply too big relative to the previous lag?


I definitely agree that larger lags are better from a convenience standpoint. The forgetting curve also shows benefit to growing gaps in the short term. But it seems more realistic to me that after a few iterations of growing lags, retention diminishew drastically. That is to say, after lags of, let’s say 1 day, 4 days, 1 week, 4 weeks, 3 months, 6 months, wider lags just don’t work becuase they are just too large. I’m not claiming this, I’m hypothesizing. Does it make sense to you that this might be the case?
After all, if you last saw something six years ago, it’s probably likely that you don’t remember it today.

This is why I’d imagine some sort of blended algorithm would be more appropriate with growing lags on the front-end and fixed lags in the later stages.

1 Like

Yes I agree, it’s only a model, and models only apply within a certain range of criterias.
I don’t know where the model stops to yield effect as the gap enlarge but I would say it’s great to see an item at least every 6 month. Maybe a year

However it raises the following statement: If you don’t see a item within 1 or 2 years, you probably don’t need it anyway and it’s ok to forget it. (From your brain perspective at least).

I see SRS as a way to familiarize with words before I encounter them in context. It helps me tremendously when I stumble upon a Kanji i learned through Wanikani.

On a side note, just like with muscle memory, there’s muscle memory with your brain. So even if you think you’ve forgotten something, it’s easier to relearn a second time. At least that’s what I felt while starting learning Japanese over and over.

I think nowadays, they speak about reinforced neural connection that helps you find the way to a particular knowledge more than forgetting something all together. So theories about pathways: the more you activate a certain pathway, the more trafic it gets and the larger the road. (In popularizing terms)

So to sum it up. SRS should be a tool not a goal,
It’s a really interesting subject though

1 Like

I agree with you across the board on this.

Something occurs to me wrt seeing items in the wild. Perhaps a solution that is closer to optimal might be only having as many stages as it makes sense to have given that a lag can be too big (so only x stages if the lag to get to x+1 would be identical as the lag from x-1 to x), after which the system you use should recommend content which you can read and which provides context to the items that you have already studied.
So when 90% of items from a particular set y are ‘known’, you are ready to consume set z of content.
A robust SRS+content recommendation system if you will, where the recommendations are appropriately timed.

This completely ignores grammar, but hey, who needs grammar anyway? :man_shrugging:

1 Like

For stopping 4 month (without studying ) and restarted to Lv 1 (from lv15 ) , i can say that the srs really work as is ,
i though i would have forgotten all but it was not the case , and i have better acuracy than before
so the srs work , and studying more can only be benificial . so i agree with that :

in fact if you study japanese ( reading , grammar ,… ) you will see word you study maybe in your lag period and the more you read the better ^^ , no need to stop reading ,… because you fear to be out of the srs system .

1 Like

I feel comfortable with Wanikani spacing. It is effective when it comes to my learning. I’m sure there are ways to smooth out the learning progression but it’s hard to find suitable content that interest people and at the same time fit perfectly their level. Sometimes I try to punch above my heights, other times I step down on a more humble content. Seems to be working so far, though only time will tell.

Most of the tool I come across recently, seemed to have a kind a SRS system built in, so everyone can find it’s calling. And maybe there will be new systems in the future fitting even more people. As long as people find value, there’s a space for a different approach

Who indeed ?
If your system includes readable content then even without dedicated grammar studying, chances are your brain will pick up pattern anyway !

1 Like

Well, my Anki deck that I’ve been working with for many years now has entries that have gaps in years that I’m doing fine with. (Concrete example, 屋外 came up this morning after an interval of 1.63 years and I answered it without trouble. I won’t see it again in Anki for another 3.69 years.)

What this really means is that I’ve seen the word often enough in the course of my general reading that it’s just a word I know. Showing it to me every six months would be wasted time. I think an SRS system must have some mechanism for getting well-known words out of your way; WK does it with the ‘burn’ idea, most others just push the interval out longer and longer. If you don’t do this, then the number of items you have to review every day just gets too big, because the total number of words available and plausible to learn is very large. For instance, if you have a vocabulary of 10,000 solidly known words and your SRS interval maximum is six months, then every day you are doing 55 reviews just to keep those 10,000 items at steady-state assuming no failures at all. That leaves much less time for learning new words.

I think a key thing to realize about vocab learning is that the aim is not to reliably remember every word you put in the system. If you put in 3000 words with a 100% recall rate that is much worse than if you put in 10000 words with an 80% recall rate. So it’s fine if a few items that got super-long intervals actually get forgotten, if the tradeoff is that you have more time each day for learning new words. And as @Makushi_Rutsu points out, an item that you never saw outside the SRS in a year is among the better candidates for not putting in the time to remember.


Definitely agree.

But then the real question becomes: why study it at all anymore? You know you’re regularly exposed to it. Your content is your SRS in this case.

1 Like

This is why just lengthening the interval exponentially is good – I don’t ever have to take the time to actively make that “I guess I know this well enough to drop it” decision, the SRS does all the work for me. And of course very occasionally I get something wrong, and then the system automatically gives me a refresher on it.

(I do make that kind of “know this well enough” judgement on the front end, when I’m putting in new words (or choosing not to). But in a review session I just want to hit the wrong/ok/easy button and move on".)


Never, either you bet that people will begin to read and encounter it in wild and drop the SRS after some arbitrary level like WK, or keep prolonging it, which will either eventually make next step so far in the future its basically as dropped, or you will start oscilating between some levels as you fail to remember a thing for that long

I havent seen any evidence that prolonging the review makes you learn it better, so yes - its not “useful” in terms of learning it better. Doing one review every week for a year (52 reviews) will no doubt make you remember it better than doing 10 reviews scaling the review length every time in the same timeframe. I dont think there is anyone really doubting that? Its just obviously very inefficient to do so if you want to learn more than handful of items.

Its true that math behind SRS cycles is a bit iffy when it gets to large intervals like years and something like gap of 6 years is truly a bet, in theory it should work, but studies on this arent that long because there simply isnt any need to SRS something for 6 years (and on top of that without seeing it anywhere else as refresher). That said its not that rare that i remember something that happened from elementary school and i havent thought about in 10+ years, so why would it be impossible with SRS item? At worst you will really reach the limit and fail it after some long period which will make the interval shorter and possible to remember, so you will oscilate on the edge of what your brain can manage, it cant get much more efficient than this, can it? (ofc this is ridiculous case in the first place, youll just encounter that stuff along the way or it was never important to learn in the first place, so no need to ever reach periods this long)


One benefit that seems overlooked, is that the exponential increase in reviews means that you can regularly add new cards without increasing your daily reviews long term (i.e. after you reach steady state); whereas if there was a fixed interval, adding new cards would increase your daily reviews. To give a concrete example, if you add 10 cards per day, eventually you’ll reach a steady state where you have to review around 100 to 200 cards per day (depending on your fail rate, how the intervals are calculated, etc.), this is because the new cards fill the gaps that are created by the exponentially increasing interval. With a fixed interval, the number of reviews will grow without bound.


This topic was automatically closed 365 days after the last reply. New replies are no longer allowed.