luke@lunalxd.com.au
Forgetting is not a failure of the learner. It is the default. This page lets you play with the two things that reliably slow it down, so you can see the shape of the problem instead of taking my word for it.
Think of one thing you learned this morning. A name, a dose, a policy change. Nothing about it was hard. Now move the slider and watch what happens to it.
One thing you learned today, with nothing done to keep it.
61%
After 12 hours, roughly 61% of it is still there.
That drop is not a personal problem. It is the ordinary behaviour of memory, and it is steep at the start and gentler later. Everything else on this page is about what you can do to the shape of that line.
Hermann Ebbinghaus measured this on himself in 1885, learning nonsense syllables and testing how many he could still produce over the following month. In 2015 Jaap Murre and Joeri Dros replicated the study and found a curve close to the one he described.
Two things set the line. How long it has been, and how well the material was learned in the first place. The second one is the lever you actually have.
Retention over thirty days, with no reviews at all.
After 7 days, roughly 25% remains.
Push the second slider up and the whole line lifts. That is worth sitting with, because most of what we argue about in course design is the first hour, and almost none of it is about what happens in the following month.
Give someone five reviews. You can spend them all at once, or spread them out. The total effort is identical. The result is not.
Each review lifts the line back to full and makes the next fall slower. Stack them on the same day and the second, third and fourth land while the material is still fresh, so they buy you almost nothing. Spread them out and each one does real work.
Set your own review days, or load a preset. Both lines run to day sixty.
Reviewing on day 1, 3, 7, 14, 30 leaves 67% at day sixty, against 0% for five reviews crammed into day one.
The sawtooth is the whole idea. Every review is a small climb back to the top, and each climb starts from a shallower slope than the last. Cramming gives you one climb and four wasted steps.
Cepeda and colleagues pulled together 839 assessments of distributed practice, drawn from 317 experiments, in their 2006 review. The spacing advantage held broadly. Their 2008 study looked at the gap itself and found that the gap producing the best retention gets longer as the interval you need to remember across gets longer.
Spacing is when you review. This is what you do during the review. Re-reading feels productive, because the material looks familiar as your eyes move over it. Having to produce it from memory feels worse and does more.
Both lines below use the same five review days. The only difference is whether the learner retrieves the material or reads it again.
The same schedule, two ways of spending the review. Both lines are drawn, and your choice is the one in focus.
Retrieving the material on those five days leaves 67% at day sixty.
The direction of that gap is well supported. Roediger and Karpicke showed in 2006 that testing beat restudying for long-term retention, and Dunlosky and colleagues rated practice testing and distributed practice as the two highest-utility techniques they reviewed in 2013. The size of the gap on this chart is not measured. I picked a factor that shows the direction without pretending to know the number.
This is the part that changes how a course is built. A knowledge check is not decoration at the end of a module. It is the mechanism.
In practice nobody schedules reviews by hand. You pick a floor you are not willing to drop below, and reviews happen whenever the prediction hits it. Choosing the floor is a design decision, and it has a price.
A review happens whenever predicted retention falls to your target, across one hundred and eighty days.
Never letting retention fall below 90% takes 21 reviews over one hundred and eighty days, about one every 8.2 days.
Drag the target from 95 down to 60 and the review count collapses. That is the trade. A higher floor costs more of the learner's time, and the cost rises faster than the floor does. The honest question in a project is not how high we can push retention, it is how much forgetting this particular content can tolerate.
For a hand hygiene reminder, quite a lot. For a drug calculation, very little.
Most mandatory training runs once a year. Here is what that pattern looks like against the same content refreshed quarterly or monthly, starting from material that was learned properly the first time.
Three patterns across 365 days. All three lines are drawn, and your choice is the one in focus.
Monthly leaves 96% at the end of the year, and no refresher has to start from scratch.
Look at the quarterly line. It never recovers, because by the time each refresher arrives the material has already fallen below the point where a review counts as a reminder. Every session is relearning from the beginning. That is the same content spread over fewer, longer sittings, and it buys much less.
The annual session is worse again. In this simulation it does not hold anything across the year, which matches what most of us already suspect about annual compliance training and rarely get to say out loud.
Three moves come out of the charts above, and none of them need a new platform.
Space it. If you have five contacts to spend, spend them across weeks rather than in one sitting. The same hours, arranged differently, produce a different result. A single long session is the most expensive way to buy retention.
Make people retrieve. Build the check so the learner has to produce the answer, not recognise it. Recognition is what re-reading gives you, and it fades faster than it feels like it should.
Aim for the right difficulty. A review that happens while the material is still fresh does very little, which is the trap in the crammed line. A review that happens after it is gone is not a review at all, which is the trap in the quarterly line. The useful moment is in between, when recall is effortful but still possible.
None of this says a learner is finished. Retention is a property of a schedule, not a score to close off. The design question is what happens in week six, not whether someone reached the end of the module.
The companion piece to this one is Beyond the smile sheet, on measuring what learners actually kept.
Everything on this page is generated by a small model running in your browser. It is not data. No measurements from any study are plotted here. The shape of the curves is the lesson and the constants are illustrative, chosen so the direction of each effect is visible.
Retention after t days is an exponential decay set by a stability value S, measured in days:
R(t) = exp(-t / S)
S is how long until retention falls to about 37 per cent. On the page I label it "how well it was learned", because that is what it means to a reader. A higher S is a flatter curve.
A successful review at retention R grows stability:
S_new = S * (1 + k * (1 - R))
with k set to 3, and the growth clamped so a single review never more than quadruples stability. Because the gain depends on 1 - R, a review at lower retention is worth more than a review taken while the material is still fresh. That is the desirable difficulty idea, and it is why the crammed schedule flattens out.
If retention at review time is below 0.3, the review is treated as relearning rather than reinforcement, and stability resets to its base value instead of growing. This is the cliff that the quarterly refresher falls off.
Passive re-reading uses the same schedule, with the stability gain scaled by 0.4. That factor is illustrative. It is set to show the direction of the testing effect without claiming to know its size.
k = 3
growth clamp = 4x
relearn floor = 0.3
re-reading factor = 0.4
base stability = 2 days (the first five sections), 30 days (the annual refresher)
A single exponential is a simplification. Real forgetting varies with the material, the learner, the sleep they got and how the memory was encoded in the first place, and a power function fits some datasets better than an exponential does. Stability here is one number standing in for a lot of separate things. The model has no concept of interference between similar items, no partial credit, and no difference between recognising something and using it under pressure. Treat the charts as a way to reason about the shape of the problem, not as a prediction about a particular person.