Anki, and spaced-repetition in general, are geared towards retaining information you’ve already learned rather than learning new information. That’s because in general, flashcards are good at reminding you of some learning experience, but not so good at creating a memorable experience to begin with.

Still, Ank provides some support for the initial learning of concepts by offering learning steps.

What are learning steps?

OK, you’re forgiven for not knowing all the details of the way Anki works. Honestly, it could be described better in the manual.

When a card is first seen, it’s placed in the learning queue. There, it goes through a series of intervals (scheduled to the minute), that are defined in the deck options. The default is two steps: 1min and then 10min.

Each time you get a review right, the next interval is the next one is the list of learning steps. You if your learning steps are 5, 10, 15, and 20, then get the card right the first time and we wait 10 minutes. Get it right again, and we wait 15; right again? 20 minutes.

Get a review wrong, and you’ll see the card again after 5 minutes (step 1) no matter where you were in the series.

If you get that last review right, the card leaves the learning queue and becomes a regular review card; Anki uses its algorithm to decide when you’ll see the card next.

Go short or go long

My thinking on this used to be that I’d only have a couple learning steps. Maybe 10 minutes and then 300 minutes. That way if I get a card wrong, I see it 10 minutes later (more or less), but if I get it right, I’ll see it in 6 hours; which means I’ll see it that afternoon, or the next morning.

I don’t intend to fire up Anki every hour on the hour, so I’m really not interested in having many learning steps that are shorter than 24 hours.

What I found doing this, was that with the more difficult cards, I could remember them for a day or so (long enough for them to become regular review cards), but once the standard algorithm set in, I’d end up forgetting them often.

That’s because the algorithm would give them intervals that were too long and it takes a while for your feedback to whittle down the ease factor enough that the intervals are appropriate.

For most cards, this worked fine; the problem was just the few difficult ones.

After a while, the few difficult cards would be marked as leeches. I have a pretty low bar for this. Only 4 lapses marks a card as as leech.

I’ve recently changed my thinking about learning steps.

Going long

I’m using much longer learning steps now. They’re equivalent to 15 minutes, 1.5 days, 7 days, and 30 days.

Yes, in order to leave the learning queue and be considered learned, I have to have remembered the card for 30 days.

These intervals came from the book How We Learn, by Benedict Carey. Apparently, these were the intervals recommended by researchers. I doubt they’re the most optimal intervals possible, but they were actually pretty close to the algorithm-generated intervals of some of my decks, so I just went with them.

Easy cards will graduate out of the learning queue pretty quickly, but harder cards will be forgotten and then bounce back to step 1 over and over until I can remember them for 30 days. While in the learning queue, lapses aren’t tallied, so cards aren’t marked as leeches. That comes after they graduate to the review queue.

This means that there’s a lot of forgetting and relearning that happens with the harder cards. In a long-running study of memory (the four Bahricks study) it was the longer two-month interval that yielded the greater long-term results; even though there was a lot of forgetting and re-learning involved. So I think these learning steps, while possibly sub-optimal, are good enough.

The four Bahrick study (named after the Bahrick family who served as the test subjects) didn’t use increasing intervals as Anki does. It used steady even intervals of days, weeks, or months. The finding was that the best long-term retention came from the longest interval; two months.

Another nice thing about using these long learning steps is that it sets sort of a bar for graduation. What I mean is, that I have to know the content pretty well (long enough to remember it for 30 days) before a card gets handed over to the scheduling algorithm. This means that I know all the cards that graduate more equally well.

What can happen with shorter learning steps is that some cards will graduate and you’ll be able to remember them for another week or so (the A students of the card world). Other cards will graduate, and you’ll just barely be able to remember them for another day (the D students). But Anki’s algorithm can’t tell the difference between these cards. It has to start off treating them all the same until your feedback tells Anki which ones are the A students, and which ones are the D students.

Setting the bar for graduation very high means that there’s much less variation in how well you know graduating cards, so it’s easier to adjust Anki’s algorithm to fit them.

Easy cards will graduate quickly, while harder cards will have to bounce around in the learning queue for much longer. The end result is that graduating cards are of a much more consistent difficulty. After that, leeches will reliably tell you which cards are poorly formatted or especially difficult and require special attention; not just more time.

At least, that’s my theory…

Personal experience so far

So far, it’s been working well. Because of these learning steps, my learning queue remains much larger than it has in the past. I also notice that the flow seems a little more efficient. Easy cards are quickly promoted and the harder cards stick around to get the attention they deserve.

As a side note, I also don’t add new cards until both the review queue and the learning queue are below a certain limit. This means, I don’t have an ever-growing backlog of difficult cards to work through, which can be a real drag. How I do this is a subject worthy of a post of its own. Maybe I’ll write about that soon.