Memorising Isn't Learning, and "Learning to Learn" Is the Skill That Actually Lasts

TheThe problem with rote learning is not that it fails to produce recall. It's that it fails to produce anything else." — a common paraphrase of the research this page is built on, not a direct quote from any one source

Ask most students what "studying" means and you'll get some version of: read the notes again, highlight the important bits, reread them closer to the exam. That's a real activity, and it produces a real feeling of familiarity. It is also, according to a large and unusually consistent body of cognitive science, one of the least effective ways to actually learn something — and the gap between "feels like it's working" and "is actually working" is exactly what this page is about.

Memorising and learning are different claims about what happened in your head

MemorisingMemorising stores a fact or procedure in a form that can be retrieved when the situation looks the same as when you learned it. Learning builds a flexible mental model — cognitive scientists call it a schema — that lets you recognise the same underlying idea when it shows up looking completely different, and lets you predict what should happen in a situation you've never actually seen before.

A concrete example from computing, since that's likely what you're studying: memorising that a for loop's syntax is for (int i = 0; i < n; i++) gets you exactly as far as writing loops that look like every loop you've already seen. Learning what iteration actually is — repeating an operation over a changing state until a condition holds — gets you to a while loop, a recursive equivalent, a functional map/reduce, and the judgement to know when a loop is the wrong tool for the job entirely. Only one of those two outcomes transfers to a problem you haven't met before.

Chi, Feltovich and Glaser's classic study of how physics novices and experts sort problems found exactly this distinction in the wild: novices grouped problems by surface features (both involve an inclined plane, so they're "the same kind of problem"), while experts grouped them by the underlying physical principle (both are conservation-of-energy problems, regardless of what the diagram looks like)¹. That's the difference between memorising and learning, observed directly: experts hadn't just stored more facts, they'd built a different, deeper kind of organisation around the same facts.

What actually builds that deeper organisation (and it isn't rereading)

Dunlosky,Dunlosky, Rawson, Marsh, Nathan and Willingham's landmark review ranked the common study techniques students actually use against the real evidence for each one's effectiveness, and the results are uncomfortable reading for anyone who studies by rereading and highlighting: both rate as low utility, despite being the two most popular techniques among students surveyed². The two techniques the evidence rates as high utility are worth knowing by name, because they're both counter-intuitive:

• Retrieval practice (the testing effect). Actively trying to recall information from memory — a practice quiz, closing the book and writing down what you remember, explaining a concept aloud without notes — produces measurably better long-term retention than the equivalent time spent rereading the same material, even though it feels harder and less fluent in the moment³. The struggle to retrieve is the learning event, not an unpleasant side effect of it.
• Distributed practice (spacing). Spreading study of the same material across several separate sessions, with real gaps between them, produces better retention than massing the same total study time into one sitting — "cramming" works for tomorrow's recognition and fails for next month's application, which is usually the actual goal.

Robert Bjork's concept of desirable difficulties ties these together into one principle: certain kinds of difficulty during learning — retrieval effort, spacing, interleaving different problem types rather than blocking them — genuinely slow down and impair performance during practice, while producing better long-term learning and transfer than the easier, more fluent-feeling alternative⁴. The word "desirable" is doing real work in that name: not all difficulty helps (confusing material badly explained is just difficult, not desirably so), but the specific difficulties above are worth deliberately choosing over the easier-feeling option.

There's a direct, practical trap hiding in the space between "feels easy" and "is working": Gick and Holyoak's research on analogical transfer found that people who'd successfully solved a problem using one underlying strategy frequently failed to apply that same strategy to a structurally identical problem in a different surface context — unless they were explicitly prompted to compare the two and notice the shared structure⁵. Knowledge that's never been deliberately compared across contexts tends to stay inert: technically present, unavailable when you need it.

Forgetting is not failure — it's part of how memory actually strengthens

ItIt feels obvious that forgetting is the enemy of learning: you studied something, now you can't recall it, so the studying "didn't work." Bjork and Bjork's new theory of disuse argues this intuition mistakes two genuinely separate properties of a memory for one another⁶:

• Storage strength — how durably something is encoded. Once genuinely learned, this barely declines, and every act of relearning it tends to ratchet it higher.
• Retrieval strength — how easily accessible it is right now. This decays quickly with disuse, and it's what "forgetting" actually is: the memory hasn't been deleted, it's become harder to reach.

The counter-intuitive part is what follows from separating the two: restudying something after retrieval strength has dropped — after you've genuinely started to forget it — builds more storage strength than restudying it while it's still fresh and easy to bring to mind. If you reread your notes the day after first learning something, retrieval strength is still high, the material feels effortless and familiar, and there's very little for that rereading to actually strengthen. Wait until you've partly forgotten it, then retrieve it with real effort (or fail to, and have to relearn it), and that effortful retrieval is precisely the event that raises storage strength. A little forgetting between sessions isn't a sign spacing has failed — it's the mechanism spacing depends on.

This has an unforgiving corollary for cramming: cramming keeps retrieval strength artificially high right up to the test by never allowing genuine forgetting to happen, which is exactly why it can produce strong performance on the test while barely touching the underlying storage strength that would let the knowledge survive a month later. It's the same "looks like progress, isn't the same claim as progress" pattern this site's project-guidance material keeps returning to — a flattering short-term number that isn't evidence of the durable thing it's being read as.

Practically: if every time you test yourself on old material it feels easy and immediate, you're probably reviewing it too soon and too often. A review that costs you some real, effortful digging — or that you partly fail — is the one doing the work.

The cognitive lies we tell ourselves about "revision"

Fittingly,Fittingly, quite a few students genuinely misread or mishear the word "revision" as "vision" — which, unintentionally, is a fair description of what a lot of revision actually is: looking at material again, passively, and mistaking the visual familiarity that builds up for actual knowledge. This is a well-documented, specific illusion, not just a turn of phrase.

Bjork, Dunlosky and Rawson's review of self-regulated learning catalogues exactly this gap between what learners believe is working and what actually is⁷. The clearest demonstration of it comes from Kornell and Bjork's study of how people learn to recognise different painters' styles: one group studied several paintings by one artist in a row before moving to the next artist (blocked practice); another group studied the same paintings in a shuffled, mixed order (interleaved practice). On a later test — classifying new, previously unseen paintings by artist — the interleaved group performed substantially better⁸. Here's the sting: when asked which method they thought had worked better, most participants said blocked practice — the one that had actually produced worse learning. Blocked practice feels more effective because each new painting is easier to place while you're still looking at several by the same artist in a row; that fluency is real, and it is not the same thing as having learned to tell the artists apart.

Rohrer and Taylor found the identical pattern with mathematics: students given a block of practice problems all requiring the same formula (solve ten volume-of-a-cone problems in a row, then ten volume-of-a-cylinder problems) scored higher on those problems immediately afterward than students given the same problems shuffled together — but on a later, mixed test requiring you to first work out which formula a given problem actually needed, the interleaved group scored far higher⁹. Blocked practice quietly gives away the hardest part of the skill for free: when every problem in front of you is the same type, you never have to practise the actual judgement of recognising which type you're looking at — you already know, because of the block structure, not because you learned to tell.

This is what "practising with different examples over a period of time" is actually doing, and it's worth being explicit that it's two separate desirable difficulties stacked together: interleaving (mixing problem types together, rather than blocking one type at a time) forces you to practise recognising which approach a new problem needs, not just executing an approach you've already been told applies; spacing (returning to the material over separated sessions, not one sitting) is what lets enough forgetting happen between attempts for each retrieval to be genuinely strengthening, per the section above. Neither one feels like it's working while you're doing it — both feel harder and less certain than the blocked, massed alternative — and that mismatch between feeling and outcome is precisely the illusion this section is named for.

In programming terms: working through ten linked-list problems in a row teaches you to execute a linked-list algorithm once you already know it's a linked-list problem — which the block structure told you for free. Mixing linked-list, tree, and graph problems together, spread across several sessions rather than one, forces you to also practise the harder, more transferable skill of recognising which structure a new, unlabelled problem actually calls for. That recognition skill — not the execution once you already know the answer — is usually the part a real exam, interview, or project is actually testing.

Why "learning to learn" specifically matters now

EveryEvery one of the ideas above has been true for decades — none of this is new science. What's changed is how much it now costs to skip it.

Domain knowledge is getting outdated faster than a typical degree lasts. This site's own AI, LLMs & Agentic Systems material makes this point directly about one field in particular: a framework, a tool, or a "best practice" learned today has a real chance of being superseded before you graduate, let alone before your career ends. Memorising the current state of a fast-moving field is studying for an exam the field itself is going to change the answer key on. What transfers is the underlying ability to learn the next thing quickly and well — which makes the skill of learning itself one of the few genuinely durable things you can study.

Generative AI has made it easier than ever to produce a correct-looking answer without the learning that's supposed to sit behind it. This site's own material on the subject documents a genuinely uncomfortable, real finding: a large panel study following over 26,000 secondary students across 30 months found that using generative AI to help complete homework improved the homework scores themselves while measurably lowering students' unassisted exam performance on the same material — the AI-assisted route produced a better-looking artefact and a worse-learned student, at the same time, from the same work¹¹. This site's Project Guidance material on scaffolding makes the mechanism explicit: an AI assistant used as a genuine scaffold — responding to your own attempt, probing your reasoning, refining something you drafted first — keeps you inside the loop where learning actually happens; the same tool used as an answer machine routes the productive struggle outside your own head and into the model's, and "the output may be excellent; the learning is zero"^{12 13}. Recognising which mode you're in, moment to moment, is now a genuinely load-bearing study skill, not a footnote.

Self-regulated learning is the actual transferable skill underneath "learning to learn." Zimmerman's widely-used model breaks it into three cyclical phases you can practise deliberately: forethought (setting a specific goal and choosing a strategy suited to it, before you start), performance (monitoring your own understanding while you work, not just executing a plan on autopilot), and self-reflection (honestly evaluating afterward what worked, what didn't, and why — then feeding that back into the next cycle's forethought phase)¹⁰. Notice that none of those three phases is about the subject matter itself — they're a skill you get better at through deliberate practice, and they transfer completely across every subject you'll ever study, unlike almost any specific fact you'll ever memorise.

A practical self-check

TheThe table below is deliberately built around signals you can actually notice in yourself, not abstract definitions.

Signal	Looks like memorising	Looks like learning
How it feels while studying	Smooth, familiar, fluent — you recognise everything as you reread it	Effortful, sometimes frustrating — retrieval fails some of the time
The test that actually distinguishes them	Can you produce the answer with the material closed, from memory alone?	Can you apply the underlying idea to a problem shaped differently from any example you studied?
What a worked example gives you	A specific answer to a specific question	A pattern you could re-derive if you'd forgotten the specific answer
What happens a month later	Recognition ("I've seen this before") without recall	Recall, and the ability to reconstruct it from principles if recall fails
The role of an AI assistant, if used	Generates the answer you needed	Reacts to an attempt you made first — checks it, probes it, refines it
Practice structure	Same problem type repeated in a block, one sitting	Different problem types mixed together, spread across sessions
How you'd rate your own progress, honestly	Confident, fluent — it feels like it's sinking in	Less certain, some retrieval attempts fail — and that's expected, not a warning sign

If you consistently land in the left column, that's worth noticing without panicking about it — it's the default outcome of studying the way most students are taught to study, not a personal failing. The fix isn't more hours; several of the changes above (testing yourself instead of rereading, spacing sessions out, comparing problems across contexts) cost the same time, arranged differently.

Where this connects to other material on this site

ThisThis page is deliberately narrow — the cognitive science of memorising versus learning, and the practical skill of learning to learn. Related material elsewhere goes deeper on specific angles this page only touches:

• Scaffolding, the ZPD, and Using GenAI Well — the full argument for treating an AI assistant as scaffolding rather than an answer machine, with a practical three-part check (fading, contingency, transfer).
• Learning as a Feedback Loop (and the AI Partner) — the same "open-loop" failure mode, argued from a control-systems angle specifically for programming practice.
• Studying Ethically: Beyond the Rulebook — why treating AI use as a rules-compliance question misses the point, and a practical AI-use continuum from tutor to proxy.

References