Patients with amnesia can follow a conversation and understand it perfectly well. They just do not remember it afterward. For Fergus Craik, one of the psychologists behind the idea that processing meaning is what makes memories last, that fact forced a concession. Understanding, he wrote in a 2002 essay, is not enough: “some further step is necessary” before an event is stored in a way that supports retrieval “minutes, days, and years later” (Craik, 2002, p. 310). He called that step consolidation, and he wrote the word with a question mark, unsure it named a single process.
Encoding, consolidation and retrieval are the three steps between meeting an idea and using it on an exam, and brain research now offers a partial bridge between those steps and the strategies in this series. It explains part of why the strategies work. It cannot plan a single study session for you, because the evidence that the strategies work still comes from behavioral studies of learners, covered in the earlier posts.
Three Stages and a Hand-Off
A review by Crowley et al. (2019) organizes memory into three stages. Encoding forms a memory in the first place. Consolidation, in Diekelmann and Born’s (2010) definition, “transforms new and initially labile [unstable] memories” into more stable ones. Retrieval uses a cue, such as a question, to bring a stored memory back to mind.
Craik’s levels-of-processing idea gets a clean test in three experiments by Peng et al. (2025). Words processed for their meaning were recognized better than words processed for their sound or their letter pattern at every delay. Both sets were then forgotten at about the same rate. Deeper encoding raised the starting point without flattening the curve. The experiments used single words, recognition tests and delays of a day or less. On these curves, the deeply encoded words stayed ahead at every delay tested, but they were lost just as fast. That puts the strategies in order: build the representation with elaboration, questioning and self-explanation, then keep it reachable with retrieval and spacing. Neither job does the other’s.
Two brain structures carry most of the consolidation story. In Diekelmann and Born’s account, the hippocampus, a curved structure deep in each temporal lobe, handles new memories for facts and events. The neocortex is the folded outer layer of the brain. In the two-stage model they describe, the hippocampus learns fast and holds new memories for a while, and the neocortex learns slowly and keeps them for the long term. During sleep, new memories are reactivated in the fast store, which acts as a “trainer” for the slow one. This reorganization is called systems consolidation.
The reactivation, called replay, was described in rats. Hippocampal cells that fired in a certain order while a rat explored a new space fired in the same order during sleep afterward. This happened mostly in the first hours after learning, and in only a minority of the recorded cells. Diekelmann and Born report signs of it in human brain imaging as well. Whether replay actually moves memories into the neocortex is a separate question. The authors write that this is assumed, “although this has not been directly demonstrated yet.” So treat the hippocampus-to-cortex hand-off as a well-supported model that is still unconfirmed.
Emotion can also change how strongly a memory is consolidated. McGaugh (2004) reviews work, mostly in rats, in which stress hormones or drugs given after training changed memory tested a day or more later. The amygdala, a small structure deep in each temporal lobe next to the hippocampus, adjusted consolidation in other memory regions. The emotional-learning post covered how far that carries over to people.
What Sleep Adds
Compared with the same amount of time awake, sleep after learning improves retention of facts and of practiced skills across many studies, with “practically no evidence for the opposite effect” (Diekelmann & Born, 2010). Benefits have been reported after a full night and after naps of one to two hours. The same review adds that the effect “is not revealed under all circumstances.”
The benefit is also selective. In one study the review describes, people learned two finger-tapping sequences and improved more overnight on the one they expected to be rewarded for at the retest. The benefit appears larger for memories that matter for later behavior and for material that was harder to encode.
Sleep can also change what a memory contains. After sleep, people who had learned single relations between objects were better at inferring a relation they had never studied. In another study, sleep helped people solve a logic problem they had failed before it, but “only if adequate encoding of the task was ensured before sleep.” All of these are laboratory tasks, and they are the only results in this post that tie consolidation to anything beyond plain retention.
How sleep does this is still debated. In the active system consolidation hypothesis, selected memories are replayed and strengthened during sleep. In the synaptic homeostasis hypothesis, waking strengthens connections throughout the brain, and sleep scales them all back down, so the connections that were strengthened most stand out against a quieter background. Diekelmann and Born think the two “are not mutually exclusive.”
One sleep manipulation sounds like a study hack. In a laboratory study by Rasch et al. (2007, as cited in Crowley et al., 2019), people learned card locations with a rose scent in the room. When the scent returned during deep sleep, memory for the locations improved. The same scent did nothing for a finger-tapping skill, and nothing when presented during other sleep stages or while awake. Crowley et al. add a caution about the related method of cueing with sounds: it has worked only in controlled settings and may not survive the real world, where similar everyday sounds compete. For now, cueing during sleep is a research tool.
The practical reading is modest. The sleep research did not test study schedules, so it cannot tell you how many days to leave between sessions. It does suggest that a night between learning and testing is not idle time, and that an all-nighter gives up a stage the brain appears to use. In the insight study, sleep helped only when the task had been learned properly beforehand, so the evening session still has to be real work. Sleep works on what you encoded; it does not do the encoding for you.
Retrieval Rebuilds Part of the Original
Endel Tulving and Donald Thomson set out the encoding specificity principle. In its broadest form, which they admitted came close to a truism, it says “that only that can be retrieved that has been stored, and that how it can be retrieved depends on how it was stored” (Tulving & Thomson, 1973, p. 359). A cue works only if its link to the memory was laid down at encoding.
Crowley et al. (2019) describe neural reinstatement, the return of encoding-related brain activity during recall, and count it among the factors that improve retrieval. They add a caution that applies to any brain study: better recall through reinstatement “does not necessarily indicate a stronger memory trace” and could reflect a better recall process instead.
Chen et al. (2017) studied reinstatement with a real story. Seventeen young adults watched a 50-minute episode of the BBC drama Sherlock in a scanner. Right afterward, they described aloud everything they remembered, with no prompts. Activity patterns for individual scenes came back during recall, including in the default mode network, a set of regions active during inward-directed thought such as remembering. The patterns were similar from one person to the next, even though each person used different words. They were also reshaped. In many regions, one person’s recall of a scene looked more like other people’s recall of it than like the brain patterns recorded while people watched it. “A memory is not a perfect replica of the original experience,” the authors write. They offer shared schemas, general templates for how a kind of event usually unfolds, as “a possible interpretation,” but did not test it.
James Nairne (2002) pushed back on the idea that matching study to test is enough. A match, he argued, is necessary but not sufficient. What counts is whether a cue singles out one memory from its competitors. A cue that fits ten similar memories equally well retrieves none of them reliably. That is an argument for practice questions that make you tell look-alike items apart: for a medical student, two drugs in the same class that differ in one side effect; in a chemistry course, two reactions that differ by a single step. These are illustrations, and they are the problem the interleaving and proactive-interference posts dealt with. Retrieval practice with feedback remains the tool, and Nairne’s argument says to make your practice cues as discriminating as the exam’s.
Learning Looks Like Shared Structure
Meshulam et al. (2021) took the scanner into a real course. They followed 20 Princeton students, all new to the subject, through an introductory computer science course in which lecture videos were watched outside class. Students were scanned five times during the semester while watching that week’s lecture videos, and again at the end while watching short recap videos and taking the final exam aloud. Five graduate-student experts were scanned for the recaps and the exam.
The study measured alignment: how closely one student’s brain activity, moment by moment, matched the typical pattern across the class or across the experts. Students whose activity during lectures more closely matched the class’s shared pattern scored higher on the final exam. During the recaps and the exam, a student’s alignment with the class was in turn correlated with alignment with the experts. A second analysis compared each student with themselves: answers that evoked more expert-like and class-like activity received higher scores. Activity from the early weeks of the course already predicted, to some extent, how students would do.
The study was small: 20 students, one course, about a seventh of the lectures scanned, and correlational findings. As the authors acknowledge, shared misunderstandings would also produce a shared class pattern, and the between-student result could reflect traits such as conscientiousness.
Cetron et al. (2019), from another group, compared 28 Dartmouth students at one point in time, half of them engineering majors. How well each person’s brain activity sorted structures such as bridges into an expert’s mechanical categories predicted their concept knowledge, but the study did not follow anyone learning. Both studies fit the same hypothesis, that successful learning involves forming representations organized like those of people who know the field.
For anyone who teaches, Meshulam et al. state the practical point themselves: “The first viewing of a course lecture, like the first reading of a textbook chapter, is just the beginning of a learning process that includes repetition and practice” (p. 7). A course built around that sentence plans for retrieval and spaced return after the lecture, so finishing the syllabus is where the work starts.
What This Looked Like for Me
I spent my first two years of medical school treating my brain like a hard drive. Read the material, write it down, read it again, store the file. When I switched to the strategies this series describes, I had no idea what was happening underneath. The new approach felt harder and worked better.
Some of my old failures now have a plausible biological story. Cramming the night before, which skipped sleep, let me down on cumulative exams. Rereading, which never tested my retrieval cues, gave me confidence that evaporated under board-style questions. The drug mechanisms I tied to real patients, which carried some emotional weight, stuck better than the ones I read off a list.
Each of those links is only a candidate. I did not run an experiment on myself, and I changed several strategies at once, so I cannot say which change did the work. The behavioral evidence told me what to do. What the biology gave me, in hindsight, was a reason not to drop the slower method when it felt worse.
The Core Lesson
Brain research fills in the middle of the story. A memory is built at encoding, reorganized over time and during sleep, and partly rebuilt when you retrieve it, so what comes back is a reshaped version rather than a replica. None of that sets a study schedule, and the reason to retrieve, space, interleave and elaborate is still the behavioral evidence.
If you keep one habit from this series, judge a study method by what you can recall a week later, whatever it felt like while you used it.
This concludes our series on the science of learning. Keep an eye out for the book it grew out of, on how knowledge is built, why it fades, and what makes it last. Follow us on Instagram @acemeded to hear when it comes out.
References
Cetron, J. S., Connolly, A. C., Diamond, S. G., May, V. V., Haxby, J. V., & Kraemer, D. J. M. (2019). Decoding individual differences in STEM learning from functional MRI data. Nature Communications, 10, Article 2027. https://doi.org/10.1038/s41467-019-10053-y
Chen, J., Leong, Y. C., Honey, C. J., Yong, C. H., Norman, K. A., & Hasson, U. (2017). Shared memories reveal shared structure in neural activity across individuals. Nature Neuroscience, 20(1), 115–125. https://doi.org/10.1038/nn.4450
Craik, F. I. M. (2002). Levels of processing: Past, present . . . and future? Memory, 10(5–6), 305–318. https://doi.org/10.1080/09658210244000135
Crowley, R., Bendor, D., & Javadi, A.-H. (2019). A review of neurobiological factors underlying the selective enhancement of memory at encoding, consolidation, and retrieval. Progress in Neurobiology, 179, Article 101615. https://doi.org/10.1016/j.pneurobio.2019.04.004
Diekelmann, S., & Born, J. (2010). The memory function of sleep. Nature Reviews Neuroscience, 11(2), 114–126. https://doi.org/10.1038/nrn2762
McGaugh, J. L. (2004). The amygdala modulates the consolidation of memories of emotionally arousing experiences. Annual Review of Neuroscience, 27, 1–28. https://doi.org/10.1146/annurev.neuro.27.070203.144157
Meshulam, M., Hasenfratz, L., Hillman, H., Liu, Y.-F., Nguyen, M., Norman, K. A., & Hasson, U. (2021). Neural alignment predicts learning outcomes in students taking an introduction to computer science course. Nature Communications, 12, Article 1922. https://doi.org/10.1038/s41467-021-22202-3
Nairne, J. S. (2002). The myth of the encoding–retrieval match. Memory, 10(5–6), 389–395. https://doi.org/10.1080/09658210244000216
Peng, N., Logie, R. H., & Della Sala, S. (2025). Effect of levels-of-processing on rates of forgetting. Memory & Cognition, 53, 692–709. https://doi.org/10.3758/s13421-024-01599-4
Tulving, E., & Thomson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, 80(5), 352–373. https://doi.org/10.1037/h0020071
