The next ten minutes of learning

You have ninety minutes, twelve topics, and an exam in eight days. You can ask for an explanation of any concept, a new set of questions, or a summary of an entire lecture.

Which request should you make first?

Producing more educational material does not resolve that decision. Even when suitable content is available, a learner must allocate limited time across acquisition, practice, review, and assessment. Memoza's central product question is how to make that allocation more useful.

A useful activity still has an opportunity cost

Imagine spending ten minutes correctly solving questions on a comfortable topic. You have practised something relevant. Yet those same ten minutes might have repaired a prerequisite that blocks three other topics, or exposed a forgotten method before the exam.

This is why the quality of an individual activity and the quality of a study sequence deserve separate evaluation. A beautifully explained solution may arrive after the learner has already understood the point. A short diagnostic question may be valuable because it changes what happens next.

The decision also depends on the goal. Preparation for a closed-book derivation exam differs from preparation for a programming assessment with documentation available. A recommendation requires some account of what independent performance will involve.

Personalisation has a research history

There is empirical precedent for using a learner's history to allocate review. Lindsey and colleagues tested personalised retrieval practice within a semester-long middle-school foreign-language course, and on a later cumulative exam personalised review outperformed time-matched massed review and a common spaced schedule (Lindsey et al., 2014). The setting concerned language learning, so it cannot settle how to sequence university proofs or clinical reasoning. It does show that review allocation can matter beyond providing the same materials to everyone.

Computational work has also explored recommending lesson sequences from student interactions. Reddy, Labutov, and Joachims modelled students, lessons, and assessments in a shared representation to support recommendations (Reddy et al., 2016). Their evaluation of prediction and sequence ranking is a different kind of evidence from a randomised demonstration of better learning. That distinction is essential when assessing an adaptive product.

A system can predict which question a student will answer correctly and still choose an unhelpful question.

Forecasting performance and improving it are different objectives.

The state a recommendation needs

For our hypothetical student, three observations could change the next action. They repeatedly fail a basic algebra step. They have never attempted one heavily emphasised topic. They answered another topic correctly yesterday with extensive hints.

These observations should not collapse into a single accuracy percentage. The algebra error suggests a possible obstacle. The unattempted topic represents missing evidence. The hinted success reveals what was possible with assistance, leaving independent performance uncertain.

A useful record would preserve the question, the response, the conditions of the attempt, and the relevant course version. It would also preserve what is unknown. If a new exam format has been announced, old practice history must be interpreted against that change.

Persistence alone is insufficient. A long record of ambiguous attempts can support poor decisions with great confidence. Data become useful when the events have interpretable meaning and the recommendations are checked against subsequent outcomes.

Information can be worth spending time on

The next activity sometimes needs to resolve uncertainty. Suppose two topics appear equally important, but one has never been tested. A short question on the unknown topic may reveal that it deserves substantial attention, or that the student can safely move on.

That diagnostic value creates a trade-off. Every minute spent measuring knowledge is a minute unavailable for another activity, although an assessment can itself provide practice. The system should spend enough time diagnosing to improve its decisions, without turning revision into an endless entrance test.

Our proposed direction for Memoza is to make these choices visible and revisable. A recommendation should give the learner a useful explanation: “Your last two attempts used hints; try this independently,” or “This topic has not been checked yet.” The student should be able to correct the system when it lacks context.

What could become a durable advantage

The business hypothesis is that reliable course models, interpretable learning histories, and validated decision policies could become valuable together. Collecting more interactions is only one ingredient. The stronger test is whether those interactions help the system allocate future study more effectively.

That advantage has to be earned. If a simple fixed rotation through the topics produces the same learning outcomes, added personalisation may not justify its complexity. If a policy saves time while preserving retention, that would be a concrete benefit worth measuring.

The next ten minutes are a small unit of product design. Repeated across a semester, they are also where a learning system spends most of the student's trust.

How Memoza fits

The record this essay asks for is the one Memoza already keeps. An attempt stores what was submitted, when it was started and finished, how many hints had been taken, and the activity it answers where there was one. The evidence it produces is filed against a concept and a course version as one of nine kinds of evidence rather than as a right or wrong flag, so a hinted success is stored as a hinted success and not as a tick. Both of those tables are append-only, enforced by a database trigger rather than by convention, and mastery is a fold over that log which can always be recomputed by replaying it. Each recommendation keeps the whole ranking it chose from rather than only the winner, the reason codes behind the choice, and which sequencing policy made it. What remains a proposal is the part the student sees: those reasons are not on the screen yet, and there is no way to tell the system that it is missing context.

References

Lindsey, R. V., Shroyer, J. D., Pashler, H., & Mozer, M. C. (2014). Improving students' long-term knowledge retention through personalized review. Psychological Science, 25(3), 639–647. Classroom study.

Reddy, S., Labutov, I., & Joachims, T. (2016). Latent skill embedding for personalized lesson sequence recommendation. arXiv:1602.07029. Computational research preprint.