# How Much to Learn with Anki

When you’re starting out with Anki or another SRS, you might wonder how much you can expect to learn with this newly efficient method of studying. Because spaced repetition follows a mathematical formula, it’s possible to do a decent job of estimating answers to these kinds of questions – certainly much better than you can manage for any other study method I know of. In this post, we’ll answer three common questions posed by new (and experienced!) users of spaced repetition systems: how much you can learn and retain in a given amount of study time, how much time you’ll need to learn a given amount of information, and how long it will take to learn and retain a single card.

**Note:** If you use the estimates in this post,
I would love to hear from you after you’ve been studying for a while
and gather more data about how well they work in practice
for a variety of people!
Please consider emailing me at `contact@sorenbjornstad.com`

.
I haven’t *made anything up* in this post to my knowledge,
but more of it than I would like is estimates and anecdotes
on top of estimates,
and the bottom estimates often have only vague theoretical backing.
Because results can vary greatly depending on
what you study and how you study it,
and spaced-repetition algorithms are based primarily on observation,
all of this stuff still involves a significant amount of
“guess and check” work.

## How much can I learn in 10 minutes a day?

The most basic question you can ask is, given a certain amount of time you’re willing to spend doing flashcards every day, how much new material should you be learning? The exact value depends on many things, like the difficulty of your flashcards and how consistently you review. But you can come up with a decent estimate; you only need two pieces of information:

**How many minutes a day can you devote to reviewing?**Remember, for best results, you want to review every day, for the rest of your life. Not the first two weeks when you’re excited about spaced repetition. Not four days a week. Not the days when you feel like it. You might want to take the number you think is good right now and cut it in half! Of course, there’s nothing illegal about getting rid of some of your cards or not using spaced repetition anymore, if you later decide you have other priorities. But the more consistent and continuous you are, the better results you’ll get, and the more realistic you are with yourself now, the more consistent and continuous you’ll be. With SRS, consistency beats volume in terms of results.**How many seconds does it take you to review each card, on average?**This usually comes to 5–10 seconds, with better cards yielding shorter times. If you have some review history behind you, you might want to look and see how long it*has*taken you. You can find this figure underneath the**Reviews**graph in Anki’s statistics if you tick the “Time” checkbox above the graph. I tend to sit just under 7 seconds. If this number exceeds 10 seconds, you definitely need to think about improving your cards.

### A simple formula

Now we can calculate how many cards we can review (*R*) in a day.
Here’s the formula,
where is in minutes and is in seconds.
(If algebra isn’t your thing,
there’s a widget a bit further down the page
that runs the numbers for you.)

As an example, suppose we want to study approximately 10 minutes per day and get through a card every 5 seconds. We get R = 108, so that’s about the number of reviews we can maintain in 10 minutes a day.

I give 54 as the constant factor in the formula instead of 60, the actual number of seconds per minute. I add a 10% penalty to your because it’s unlikely you’ll actually be studying cards for 100% of the time you’ve set aside to study: on a typical day, you’ll probably stop to fix one of your cards, or look something up, or go make a cup of tea. If you’re bad at concentrating in the digital age – like most of us! – you might want to reduce the number a bit further.

**Warning**:
If you want to check this estimate against how long you actually spend studying,
be aware that Anki won’t always include your unproductive time in your statistics –
if you aren’t looking at a card, it’s not counting time,
and if you spend time looking at a card, go to another screen, and come back,
the time you spent before going elsewhere won’t be counted.
Similarly, if you undo a review (because you chose the wrong rating)
and pick again,
the time you spent before undoing won’t be counted.
That’s why this time should be considered separately.
If Anki says you studied for 15 minutes,
it might have taken something like 17 minutes of clock time.

A commonly bandied-about rule of thumb is that your reviews will, over time, mount to roughly 10 times your new cards per day. I’ve found this to be tolerably accurate. So our figure of 108 means we can afford to add 10-11 new cards per day if we don’t want to exceed 10 minutes of study.

Here’s a little JavaScript widget that does the math described above for you:

### Implementation details

If 10 new cards per day sounds small, it is and it isn’t.
If we assume that you create two to three cards
about every useful thing you learn,
you’re durably learning about 4 carefully selected things every day.
(If you’re not sure why you’d create more than one card,
be sure to read up on creating precise cards
and the Minimum Information Principle.)
You can get quite a lot of mileage out of that;
after all, most people don’t select *any* items to learn every day,
they just remember whatever they happen to remember.
To put it another way,
after a year of learning 4 new things per day,
you’ll know about 1,500 new things of your choosing –
not just any 1,500 things,
1,500 things that you know will be useful in your daily life
and that you won’t forget.
Of course, if you can study for 20 minutes every day,
you’ll learn twice as much, which is great…
but there isn’t really a minimum amount of useful study.
Consistency beats volume.

Now, just because you can only add 10 new cards per day
doesn’t mean you literally have to click the “add” button
and type in exactly 10 cards each day.
You can add them in whenever you think of them
or have time to batch-add some content;
your SRS will take care of queueing them up
and only introducing the number you ask for every day
to keep your workload steady.
(If you end up with a large backlog of new cards,
any good SRS will offer tools to reprioritize them.)
To change the number of cards introduced per day in Anki,
visit your deck’s options (gear icon in the deck list),
click the **New Cards** tab,
and adjust the number of “New cards/day”.

Anki doesn’t introduce new cards if you miss a day of studying – for example, if you miss 5 days of study and you’ve chosen to add 10 new cards per day, on the day you come back you will only get 10 new cards, not 60. If you’re trying to get through a large number of cards in a certain amount of time and expect to miss a few days here and there, it’s important to take this into account. Of course, you can always choose to add extra new cards in, but Anki won’t do that unless you tell it to.

Lastly, don’t forget to consider the time it takes to *create* cards.
This is usually small compared to your review time,
but not insignificant –
it’s reasonable to imagine it might take 15 seconds to a minute
to create each card,
depending on what you’re studying and how much experience you have.
You’ll also have to spend some time editing and maintaining the cards later.
If you study 10 minutes per day,
and you’re creating all your own cards,
you might end up spending more like 12 to 15 minutes a day
in Anki.

### Rule of thumb: 1 for 1

You might notice that, in our example,
we said 10 new cards per day would require 10 minutes of daily review.
With the simple formula above, it’s reasonable to extrapolate the estimate
and assume that 20 new cards per day will take 20 minutes,
and so on.
In general, over the long term,
*one new card per day equals one minute of daily review*.

Of course, this is a rule of thumb on top of a rule of thumb, so expect it to be wildly off on occasion. In particular, the formula/widget will give a noticeably better estimate if your cards take longer than 5 seconds to study or your productivity value is much off of 90%. Nevertheless, this rule is wonderfully easy and not totally imprecise for an off-the-cuff estimate.

## How much work will it be to prepare for an exam?

That covers how many cards you should study if you have a specific amount of time to spend. But in some circumstances, you may want to know the opposite: given a certain amount of material you have to master before a certain time, how much new material should you be introducing and how much time is it going to take to learn it? The following calculator does its best to answer those questions without any knowledge of your material.

The formula here is merely a combination of the one above and some obvious dimensional analysis (e.g., the number of cards you have to learn is equal to the number of facts you have to learn times the number of cards per fact), so I won’t go through it in detail.

Important notes about the meaning of the variables here:

- I separated “cards per fact” and “facts to learn” for convenience if you don’t have all the cards created ahead of time. Perhaps you have a question bank of 1,500 facts you want to learn, and in the initial set of cards you’ve created, you averaged 2.5 Anki cards per question. (If you’re not sure why you might need to create multiple cards per question, be sure to read up on creating precise cards and the Minimum Information Principle. If you don’t do that, you’ll likely be spending more time studying than estimated here.) If you already have the cards created, just leave this at 1 and fill the number of cards in for “facts to learn.”
- Reducing the number of “study days per week” from 7
will mean no new cards are introduced
on the days you do not study,
and the widget will account for that.
Your study time will be much higher the day after you skip one or more days,
since some cards will be overdue;
the “converged study time/day” field
*does not*average this in, it continues to list the estimated amount of time you will spend reviewing on those days when you don’t have any overdue cards. In my experience, you can expect to spend a wee bit less than twice the time the day after you skip. - The word “converged” in “converged reviews/day” and “converged study time/day” reflects the fact that it will take some time after you begin studying – probably several weeks – to reach this level of study time. In the days prior, it will be lower. You can attempt to flatten this out by introducing more cards than “new cards per day” at the start, but be careful about overshooting: the due reviews can mount up fast! You absolutely don’t want to add as many new cards as it takes to fill out a normal study session, or at least not for long; if you add 150 for just three days, suddenly you’ll have 450 cards due for review within a five-day period or so.
- The
*new cards per day*figure is exact (well, it’s rounded to two decimal places) and a matter of basic arithmetic; if you study at least this number of new cards per day, as long as your other figures are correct, particularly the ones about how many cards you end up creating, you are guaranteed to get through all of the material before the exam. The*converged reviews/day*and*converged study time/day*are estimates based on the model explained earlier in this post, resting on the 10x rule of thumb. - The calculator does not take into account the time you’ll be spending initially reading and understanding the material (if you haven’t done so already), creating cards (if you’re creating your own), and editing cards (which you’ll want to do regardless).
- The calculator will take you right up until the exam, so you’ll still be learning new material the day before. It’s advisable to add an additional week or so, in case you study fewer days than you expect, and so you have a little bit of time to consolidate your knowledge and perhaps take a practice test or two.
- On that note, and most importantly, since this is an
*estimate*, it’s wise to build in a margin of error and periodically review your progress so you can adjust if it proves to be low, especially if you haven’t created any cards yet and don’t necessarily know how many will be needed to master the material. Even if you’ve done this before, every topic is different. I am not responsible if you use this calculator and aren’t ready for your exam!

## How long will it take to learn one card?

A final useful thing to quantify is how long it takes to learn one card and maintain it for the remainder of your lifetime. This might help you decide whether it’s worth learning a particular piece of information, for instance. The figures should be the same for someone of just about any age, as most of the effort happens early on. If you’re at or above retirement age, you might be able to subtract 10%.

**tl;dr**: 2-5 minutes,
more likely on the low side,
depending on how optimistic,
determined to learn everything,
and careful at creating flashcards you are.

Cards vary wildly in how much review time they require. Leeches might take 20 minutes of your time and still leave you ignorant; the easiest cards with straight easy ratings might take 30 seconds over your entire lifetime. (On the default scheduling settings, a card with straight easy ratings progresses through the sequence 4 days – 13 days – 1.5 months – 5.4 months – 1.7 years – 6.9 years – 29 years – 100 years; at 5 seconds per review, that’s 40 seconds total. The last interval would be 128 years, but Anki caps intervals at 100 years by default.)

However, people come up with wildly different *average* estimates as well;
for example:

- SuperMemo’s theory page suggests you can learn 200–300 items/year/minute (i.e., if you expend one minute studying every day for a year, you’ll acquire 200-300 items, with their lifetime review costs included). If you run a couple of conversions on that, it comes to 1.2–1.9 minutes per item.
- Gwern comes to a similar conclusion of 1.8 minutes, explaining how this figure can be derived using a more complex formula on the same SuperMemo page, but then decides to more than double the number to 5 minutes, thinking SuperMemo’s model may prove too optimistic in real-world use.
- Michael Nielsen finds Gwern’s estimate too optimistic in turn and opts for 10 minutes, using an ad-hoc but sensible-feeling reasoning process and his short-term review history.

That’s a full range of 1.2–10 minutes, spanning an entire order of magnitude! Who do we believe?

I’ve been using Anki for 10 years, albeit sometimes on and off. Let’s take a look at how these estimates have held up in practice for me. I dug into my collection and took a look at those cards that have an interval of greater than 1 year (n = 18541, mean interval = 6.01 years). Cards that have reached 1 year have generally accumulated a majority of their lifetime review cost already because intervals quickly rise exponentially beyond a human lifetime from that point. Here’s a quick summary of the current total review times on these cards. Times are in minutes:

**Mean**: 0.944**Median**: 0.561**Q1–Q3**: 0.30–1.14

You might notice just from the difference between the mean and median that the data are strongly right-skewed – that is, the extreme values are all on the top end. If you’re into statistics, here’s a nice box plot showing that skew:

Here’s another neat way to look at it. This shows the current interval of each card versus how much total time I’ve spent reviewing it, but colored by the ease (the ease values are shown in per mille, or parts per thousand – 10 times the normal percentage value):

We can see that both the total review time and the interval
cluster towards lower values,
and also that the more difficult cards tend to sit at lower intervals
*and* higher review times.
The tail to the upper-left visually demonstrates the benefit of
trashing the most difficult material in your collection –
and this graph is after I’ve already trashed the very worst from mine;
an untrimmed collection would be even worse!

At any rate, we can see that the majority of my cards have accumulated less than 1 minute of total review time after attaining an interval of at least a year. As mentioned earlier, the period of 1 year is significant because at this point, unless you forget the card, only a handful of repetitions remain in your lifetime, less than half of the total repetitions for the card.

But how many cards do we forget? If we forget one, we have to start the progression over from the beginning, or at least close to it. Here’s the summary for lapses:

**Mean**: 0.82**Median**: 0.00**Q1–Q3**: 0.00–1.00

In other words, fully half of my cards have never lapsed at all, even with interval of over a year. 75% of them have lapsed once or never, and even taking into account the outliers (some of which have absurd values like 16), the average card lapses less than once. I see no reason to believe that this number will increase much further for the oldest cards as time goes on; others have been doing spaced repetition for long enough that if the spacing effect broke down at that point, we would know about it by now. Further, assuming the spacing effect does hold out over the lifetime of a card, chances are that most lapses occur in the earlier stages, because most reviews do; that means fewer cards are likely to be forgotten past an interval of 1 year, where the most rework is required.

(I haven’t taken a rigorous look at it, but I think this is how Nielsen gets so far afield from the other estimates: he uses the average calendar time between lapses per card to quantify how often cards will lapse and have to be relearned from the start. This doesn’t smell right to me, simply because there are far fewer reviews to lapse on when the intervals get high. Since the algorithm aims to create an equal chance of forgetting the card for every review, if you imagine 1 year is a bit more than halfway through the review schedule with 7 reviews before and 5 after, by the time you hit 1 year, your chance of ever lapsing again is already less than over the previous 7 reviews, even though there may be 30 times the amount of calendar time to do it in. This would seem to result in an overestimate of the average lifetime lapses per card. Nielsen also says he rarely uses the “hard” or “easy” ratings in Anki, which could reduce his efficiency somewhat, but I have no data on how much efficiency this leaves on the table.)

Seeing this data, I feel pretty comfortable saying my average lifetime cost is going to be somewhere in the neighborhood of the value predicted by SuperMemo. 2 minutes seems like an entirely reasonable estimate. 3 minutes is a nice upper bound even if I missed a few things in this analysis. 5 minutes is conservative and an easy round number, so it’s not a bad rule of thumb when considering if it’s worth learning some particular fact, as Gwern uses it – but it’s high enough that it might lead you astray if you’re trying to decide whether it’s worth learning, say, 20,000 cards.

Now, as alluded to earlier, I do practice the removal of highly difficult material in my collection, though it probably amounts to 1 or 2 percent, nothing like the 10 percent SuperMemo theory considers. Reviews for cards that are no longer in my collection were not considered in my dataset. That means, first of all, that my average review time is somewhat greater than what I’ve quoted, since I’ve spent time reviewing harder-than-average cards I later deleted. I think there’s still enough margin between my numbers and the SuperMemo numbers that this won’t affect my analysis. Secondly, it means that if you insist on keeping all material, even the most difficult and least important, you may come up noticeably worse than me (remember that tail in the upper-left of the graph!).

Similarly, if you don’t write good flashcards, more of your cards could go in the direction of 5 minutes. But it’s worth noting I haven’t been religiously following my Rules for Designing Precise Anki Cards for all of the past ten years. I’ve created my fair share of mediocre cards, and many are in there still. So my performance isn’t hopelessly skewed away from what you could hope to get because I’m a Spaced Repetition Expert.™

I also certainly have some cards that are very easy in my collection, which would perhaps lower the average lapses and review time a bit, but this is a good thing that you should do too: easy cards cost hardly any time to review, while the cost of forgetting something “easy” is high.

**A more technical concern**:
I have a recent long hiatus in those 10 years of statistics,
for several years after I got out of college,
and I haven’t yet caught up on all of the overdue reviews,
which means it’s reasonable to suppose
I might be somewhat underestimating the review time
since I haven’t yet told Anki which of those cards I’ve forgotten
and started the process of relearning them, costing additional time.
This said, I took a quick look at the cards I have caught up on
and those I haven’t
and found that the cards I haven’t
show up consistently easier on every measure than those I have,
including measures which definitely wouldn’t be affected by being behind,
like the average review time per card.
Also, even with reviews delayed by 3 years in some places,
I’m still recovering 50-75% of the memories, depending on topic,
which is much less than the 90% promised by on-time reviews
but not bad at all given the circumstances.
That means the cards I remember get an extra boost
because they were obviously easier than their statistics would have suggested
(else Anki would have scheduled them further in the future in the first place),
which helps to counter the extra review time for the ones I’ve forgotten.
So, overall, I think this effect is going to be weaker
than I would have guessed.
If my results change significantly over the next couple of years,
I’ll be sure to make an update.

**Scheduling note**:
About halfway through my spaced-repetition career,
I switched Anki’s scheduling algorithm
to reduce lapsed cards to 10% of their former interval
rather than the default of 0%.
I haven’t done any rigorous research into the effects of this,
but my impression is that it reduces the pain of failing a card
without making it substantially more likely to be forgotten again immediately.
Damien Elmes, the main developer of the Anki algorithm,
has said in the past that the only reason he put the default at 0%
is that it’s less confusing for new users that way.
I include this bit only because there’s a chance it could make my learning
slightly more (or less!) efficient than others’.

### Using your own data

If you have a bit of shell and data analysis know-how and want to play with this data from your own collection, here’s how to get the data out. This assumes a Linux-like or MacOS system, but you could adapt it for Windows as well.

Place the following SQL query in a file called `query.sql`

:

```
SELECT
CAST(SUM(revlog.time) AS FLOAT)/60000 as t,
notes.id,
notes.tags,
decks.name,
cards.reps,
cards.lapses,
cards.factor,
cards.ivl
FROM revlog
INNER JOIN cards
ON cards.id = revlog.cid
INNER JOIN notes
ON notes.id = cards.nid
INNER JOIN decks
ON cards.did = decks.id
WHERE cards.ivl > 365 -- 1 year
GROUP BY cid
ORDER BY t DESC;
```

Then make sure Anki is closed (or your collection will be locked and unavailable for querying) and use the following shell one-liner to create a pipe-separated values file:

```
sqlite3 "~/.local/share/Anki2/Soren Bjornstad/collection.anki2" < query.sql | awk -F '|' 'BEGIN { OFS = FS; print "ReviewTime|NoteId|Tags|Deck|Reps|Lapses|Ease|Interval"; } { gsub(/^_/, "::", $4); print }' > records.psv
```

For the `^_`

after `gsub`

,
that is a literal “unit separator” ASCII character –
to type it, in your terminal press Ctrl+V, then Ctrl+underscore.

You’ll need to have SQLite installed,
and you’ll want to replace `Soren Bjornstad`

with the name of your Anki profile
(if you don’t know what a profile is, yours is probably called `User 1`

).
If you’re on a Mac, you’ll have to change the `~/.local/share/Anki2`

bit too
– see file locations
in the manual.
Once you get the one-liner to run,
just import the `records.psv`

file
into your favorite analysis software or spreadsheet.