First, some context
In the late 19th century, a psychologist named Hermann Ebbinghaus spent years understanding the behavior of human memory. He charted the rate of memory decay and summarized his research with something called the Forgetting Curve.
Ebbinghaus showed that forgetting occurs rapidly in the first few hours or days after learning, and then levels off gradually over time. The rate of forgetting is steepest at the beginning and then becomes less steep as time goes on. The two factors that influence the slope of the forgetting curve are:
Storage Strength: the durability and stability of a memory over time. It is the measure of how securely a memory is initially encoded and stored in the brain.
Retrieval Strength: the measure of how easily a memory can be retrieved from long-term storage and brought to conscious awareness.
One helpful analogy here is to imagine a large library. Books are kept on shelves throughout the building. That’s storage. One’s ability to locate a book, however, depends on the accuracy of the card catalogue (and one’s prowess at way-finding). That’s access.
Robert Bjork of UCLA shows that accessing a given memory improves our ability to access that memory later. Ironically, access improvement is inversely related to how much access decay has occurred. In other words, when a memory is accessed just before it’s forgotten, we see a stronger improvement in our ability to access that memory later on. And it’s the repetition of this activity — accessing just before forgetting — that gives strength to spaced repetition as a learning vehicle.
Repeated memory accesses and at expanding intervals have the effect of flattening the Forgetting Curve. And this process makes it much easier for us to retain new information.
How Yak Tack used to do things
Enter spaced repetition, an approach that borrows from this research. Spaced repetition is the center of Yak Tack. And we built our first algorithm on the Fibonacci Sequence, which starts as 1, 2, 3, 5, 8, 13, 21, and 34. We applied it statically, meaning we’d move each word through the sequence on schedule, no matter what.
While effective, there are myriad factors our Fibonacci algorithm ignored. What if someone doesn’t read their tack notifications in time, only to find a pile of notifications for intervals 1, 2, and 3 on their Tack Board? What if someone has a hard time remembering certain words?
We knew we needed to find a more adaptive algorithm that would tune itself according these situations.
How we do it now; the Leitner system
We took the opportunity to build a new spaced-repetition algorithm based on the Leitner system. It was developed by Sebastian Leitner in the 1970’s. And while it’s a simple system, it gives us the adaptation we were looking for. We reduced the tack duration from 34 to 13 days. And we test people more frequently than we used to. With each successful test, the repetition interval expands. With each unsuccessful test, we tighten the interval.
Here’s how it works on Yak Tack for someone we’ll call “Ann”:
Ann tacks the word oligarchy. Yak Tack starts the word at proficiency level 1 (L1) for her. At this point, Anne’s ability to remember the word is decaying fast.
Ann reads her tack notification for oligarchy the next day. We promote her to proficiency level 2 (L2).
She reads her notification again on the second day. Now she’s promoted to level 3 (L3).
On the third day, we ask Ann to pick the right definition for oligarchy. She makes the right choice, now she’s promoted to level 5 (L5).
On the fifth day, we ask Ann once again to pick the right definition. This time, she chooses the definition for the word monarchy, not oligarchy. We demote her back to level 3 (L3) for oligarchy.
On the seventh day (5 - 3), we ask her to pick the definition again. She gets it right this time, and she’s once again promoted to level 5. She’ll receive the next test two days later.
This repeats until Ann reaches level 13, at which point her Forgetting Curve for the word oligarchy should be sufficiently flattened.
If words are remembered at every interval, one can flatten the Forgetting Curve in about 13 days. It’ll take a bit longer if mistakes are made along the way (which we expect!).