How spaced repetition changed the way I design drills.

The first version of TradeInTune's drill engine was bad. Not bad in the way bugs are bad, bad in the way the underlying assumption was bad. The assumption was: if a learner gets a scenario wrong, show it to them again immediately, so they can lock in the correct answer.

Every learning intuition I had said this was right. Every parent and every teacher I had ever met agreed. You got it wrong, here is the correct answer, do it again so it sticks. Within the platform's first month of testing, retention numbers told me the assumption was wrong.

The retention test was simple. Show learners a setup on day one. Ask them to identify the entry. If they got it wrong, surface it again on day two. Then test them on day fourteen and ask them to identify the same setup type, on a different chart. The "show it again immediately" group performed slightly better on day two. They performed worse than a control group on day fourteen. Some of them performed substantially worse.

The cognitive science explained it. The phenomenon is called the spacing effect, first measured by Hermann Ebbinghaus in 1885, and the most replicated finding in experimental psychology over the last hundred years is that information learned with gaps between exposures is retained more durably than information learned in a single concentrated session. A 2006 meta-analysis by Cepeda and colleagues, pulling together three hundred and seventeen experiments, quantified the optimal gap as a function of the desired retention interval. To retain something for a week, space the reviews about a day apart. For a month, one to two weeks. For a year, several weeks.

What the research said was that retrieval which is on the verge of being forgotten is what strengthens the memory trace. Easy retrievals do not build durable memory. Effortful ones do. Showing the same scenario twice on the same day produces a series of easy retrievals, because the answer is fresh, and easy retrievals do not produce durable encoding. The version of the learner who answered the same question twice on the same day felt more confident on day two and was more confused on day fourteen, because they had built a fluency illusion rather than a memory.

So the drill engine was rewritten. The new version of the engine is described informally as a forgetting-curve scheduler, though the underlying algorithm is closer to a Bayesian model of per-concept knowledge, with intervals tuned against the rough Cepeda spacing parameters. When you get a setup right, it goes back into the queue with a lengthened interval. When you get it wrong, it goes back into the queue with a shortened interval, but never the same day. The next time you see it is tomorrow at the earliest. By the time you see it again, you have had to forget enough of the easy retrieval to make the next retrieval effortful, and the effort is what builds the durable memory.

This is also why the drill set composition matters. If today's session is five drills on the same concept, the second through fifth drills are all easy retrievals, because the concept is loaded into working memory from the first drill. So the queue interleaves. Five drills, five different concept families, with the harder ones biased toward areas the learner has been weaker on. Bjork called this "interleaved practice" in 1994. It feels harder. It produces transfer that blocked practice does not.

The adaptive layer sits on top. Each learner has a per-concept decay rate the model is estimating in real time. Risk-sizing concepts decay differently from chart-structure concepts, and they decay differently for different learners. The scheduler is therefore not a single curve. It is roughly forty curves per learner, one per concept family, each independently estimated from accuracy and reaction-time signals.

The first month of running this revised engine, retention numbers across the platform improved measurably. Learners who would, under the old engine, have plateaued in week three were still improving in week six. The learners themselves did not feel like the platform had changed. The drills felt about the same. The difference was structural and invisible: the schedule was right.

There is a reason you cannot tell when good spaced repetition is happening. If you can feel it, the spacing is wrong.