In 1972 Fergus Craik and Robert Lockhart published a short, awkward paper arguing that psychology had the memory question backwards. The field was busy modelling memory as a set of stores — sensory, short-term, long-term — with information passing between them. Craik and Lockhart proposed that the stores were less important than the processing: what survives is a by-product of how deeply you thought about something at the time.
The framework has three rough levels:
- Structural — what it looks like. Is the word in capitals?
- Phonemic — what it sounds like. Does it rhyme with train?
- Semantic — what it means. Does it fit in the sentence "He met a ___ in the street"?
Ask people one of those questions about a word, then surprise them with a memory test. Semantic beats phonemic beats structural, reliably, even though nobody was trying to remember anything. In Craik and Tulving's 1975 follow-up the semantic advantage was large — and the participants had no idea they were in a memory experiment at all.
Why this matters more than it sounds
The practical consequence is blunt: intention to remember is almost worthless, and depth of engagement is almost everything.
You can decide very hard to remember a phone number and fail. You can read one sentence that reframes an argument you care about and never lose it. The difference is not effort or motivation, it is what your mind did with the material in the seconds it was there.
That explains a set of findings that otherwise look unrelated:
- Re-reading is shallow. The second pass processes the same surface at the same depth, which is why it feels productive and does so little. It also inflates the illusion of competence.
- Highlighting is structural, or close to it. Deciding a sentence looks important is not the same as deciding what it means.
- Explaining something in your own words is semantic by force, which is most of why the Feynman technique works.
- Asking why is this true? is semantic processing with a prompt attached — see elaborative interrogation.
- Producing an answer rather than reading one gets you both depth and retrieval, which is the generation effect.
The problem the theory never solved
Levels of processing has an obvious hole, and it was pointed out almost immediately: depth is defined by what gets remembered, and what gets remembered is explained by depth. There is no independent ruler for how deep a given operation is. If a semantic task produces poor recall, the theory can retreat to saying the processing was not really that deep.
That circularity is why it is called a framework rather than a theory, and why the field moved on to more specific mechanisms — elaboration, distinctiveness, and transfer-appropriate processing, which showed that "deep" is not universally better but better for the kind of test you will face. Process a word for its sound and you will do relatively well on a rhyming test. The match between how you encode and how you will be asked matters, which is the same insight running through context-dependent memory.
None of that makes the original wrong in the way that matters to you. Semantic processing beats surface processing for essentially every test you care about in real life, because real life asks about meaning.
What to do with it when taking notes
The whole framework collapses into one question: did anything happen to this material in my head before it hit the page?
- Never copy a sentence you can compress. Copying is transcription and transcription is shallow. Rewriting it in eleven words means you had to understand it. This is the mechanism behind progressive summarization.
- Write the note as a claim, not a topic. "Pricing" is a label. "Raising the price filtered out the customers who generated the support load" is a thought, and only one of those two survives a month. That is the point of atomic notes.
- Connect it to something you already know. Every link you make deliberately adds a retrieval route. This is why the linking habit in a Zettelkasten works even when you never re-read the linked note.
- Prefer questions to headings. A heading is structural; a question is semantic and forces an answer. It is the mechanic driving the Cornell method's cue column.
- Do not confuse effort with depth. Beautiful handwriting, colour-coding and elaborate templates feel effortful and are mostly structural. Ten seconds of asking so what? beats twenty minutes of formatting.
The uncomfortable corollary for anyone who takes a lot of notes: a note you wrote without thinking is only a search result. It is worth having — external memory is a real upgrade — but it did not teach you anything on the way in. The depth has to be paid for somewhere, either when you write it or when you review it. More on the machinery in how memory works and the wider section on memory and focus.