How memory works: the pipeline, and which technique acts on which stage¶

Level: 101 · for anyone who wants to know why the study techniques work, not just that they do

One line: A memory passes through attention, working memory, consolidation and long-term storage, and each stage has a limit; every technique with strong evidence acts on one of those limits. The techniques are well established; the biology named under each stage is mostly good evidence in animals.

The pipeline¶

attention / encoding  ──>  working memory  ──>  consolidation  ──>  long-term storage
        │                        │                     │                     │
  acetylcholine             bottleneck            sleep / rest           re-retrieval
  (a focused beam)       (3–5 chunks at once)    (LTP and BDNF)     (the trace strengthens)

The top row is a model every textbook uses, from Atkinson and Shiffrin's multi-store model (1968) through Baddeley and Hitch's working memory (1974). The bottom row is the biology usually given for each stage, graded on brain chemistry. What this page adds is the column that matters to a learner: what limits each stage, and what to do about it.

Stage	What limits it	What to do	Grade	Lesson
attention / encoding	what you did not attend to was never encoded	remove extraneous load; know the goal before you read	well established	cognitive load, count the vowels
working memory	about four chunks at once (Cowan, 2001)	make the foundations automatic, so each takes one place	well established	cognitive load
consolidation	a new trace is weak, and the next thing you learn can overwrite it	review within the day; sleep instead of cramming	well established	spaced retrieval, sleep
long-term storage	a trace fades unless it is used	retrieve it, spaced out and mixed with other topics	well established	spaced retrieval, interleaving

And beside every stage, metacognition: a test, not a feeling, says whether the trace is there.

Three processes¶

Zakrajsek's The New Science of Learning (chapter 4) reduces the memory models to three processes, after Melton (1963):

Encoding is taking the information in. There is a kind for every sense; most of what a course teaches is encoded semantically, as meaning, and elaboratively, by attaching it to what you already know.
Consolidation stabilises the trace. It happens in two phases: within hours at the synapse, and over days to months as the memory is reorganised from the hippocampus to the cortex (Sun and colleagues, 2023, doi:10.1038/s41593-023-01382-9 ↗, model the second phase as complementary learning systems). Sleep does much of this work: sleep. A new trace is easily interfered with: meet Jordan and then Pablo seconds apart and you keep Pablo. Hence the advice to avoid back-to-back classes where you can, and to review a class before the next one starts. Grade: well established for the effect, good evidence for the two-phase account.
Retrieval gets it back out, and is not neutral: each retrieval destabilises and restabilises the trace, reconsolidation, usually stronger (Nader and Hardt, 2009; in rats, Gonzalez and colleagues, 2021, doi:10.1073/pnas.2025275118 ↗, show dopamine gating whether new information updates the reactivated memory). This is why testing yourself is a study method and not just a measurement. Grade: well established in people for the testing effect; the reconsolidation mechanism is good evidence, mostly in animals.

The forgetting curve¶

Ebbinghaus (1885) measured his own memory for nonsense syllables and found the shape every study since has confirmed: loss is fast at first and then slow. Roughly half of new material is gone within a day or two and most of it within a month, unless it is reviewed (DeSoto and Roediger, 2019, is one modern replication). Zakrajsek's own Calculus 1 story is the cost in one number: high Bs and low As on four unit exams studied all night, then 37% on the final. Spaced retrieval has the arithmetic of why reviews at growing intervals keep a fact for years at a cost that grows only with the logarithm of the time. Grade: well established.

The big four, and how sure each claim is¶

Summaries of learning science, including AI-written ones, tend to name the same four techniques. Three have lessons here.

Technique	What to do	Grade	Lesson
retrieval practice (active recall, the testing effect)	close the book and write down what you remember; answer questions before checking notes	well established: Agarwal, Nunes and Blunt's review of classroom studies (2021, doi:10.1007/s10648-021-09595-9 ↗); Van Hoof and colleagues (2021, doi:10.1097/CEH.0000000000000335 ↗)	metacognition, spaced retrieval
spaced (distributed) practice	review at growing intervals; let Anki schedule it	well established: Cepeda and colleagues (2006), 259 of 271 comparisons; Hintzman (1974) for the theory; Van Hoof and colleagues (2021, doi:10.1097/CEH.0000000000000315 ↗)	spaced retrieval
interleaving	mix topics and problem types in one session	well established for problems that are easy to confuse; Chen, Paas and Sweller (2021, doi:10.1007/s10648-021-09613-w ↗) argue it works by discriminative contrast, a different mechanism from spacing	interleaving
elaboration (elaborative interrogation, dual coding)	ask why; connect it to what you know, in your own words and in a drawing; explain it as if teaching	good evidence: Dunlosky's 2013 review rated elaborative interrogation moderate; Van Hoof and colleagues (2024, doi:10.1097/ceh.0000000000000580 ↗)	studying vs learning: teach-the-material mode

Donoghue and Hattie's meta-analysis of ten techniques (2021, doi:10.3389/feduc.2021.581216 ↗) and Karpicke and O'Day's handbook chapter (2024, doi:10.1093/oxfordhb/9780190917982.013.70 ↗) reach the same ranking. Two caveats belong beside the top row. Retrieval practice is effortful: Zheng, Sun and Liu (2023, doi:10.1038/s41539-023-00159-w ↗) found its benefit depends on having working-memory capacity to spare, which is a reason to test yourself rested and in a quiet room, not a reason to skip it. And students can mostly recognise the effective techniques and still not use them (Rea and colleagues, 2022, doi:10.3390/jintelligence10040127 ↗); what moves them is self-efficacy and habit, which is mindset and motivation and goals and schedules.

Desirable difficulties¶

Robert Bjork's phrase for what the big four have in common: conditions that make learning harder now and better later. Retrieval, spacing and interleaving all lower performance during practice and raise it on a later test, which is why learners misjudge them (Nelson and Eliasz, 2022, doi:10.1111/medu.14916 ↗, for a review; Biwer and colleagues, 2020, doi:10.1016/j.jarmac.2020.03.004 ↗, on teaching students to choose them anyway). Two related findings, often garbled in summaries:

Pretesting. Guessing the answer before you study helps, even when the guess is wrong (Kornell, Hays and Bjork, 2009; Richland and colleagues, 2009). Errors are useful feedback, and larger errors, felt as well as measured, predict a larger testing effect (Shi and Liu, 2026, doi:10.3758/s13423-026-03009-z ↗). Feedback after a hard retrieval matters: Kliegl, Bjork and Bäuml (2019, doi:10.3389/fpsyg.2019.01863 ↗) show it can reverse the usual cost of effortful retrieval, and hints can rescue a retrieval that would otherwise fail (McLane and Selmeczy, 2025, doi:10.1080/09658211.2024.2406312 ↗). Grade: good evidence.
Hypercorrection is a different finding: errors made with high confidence are the easiest to correct once feedback arrives (Butterfield and Metcalfe, 2001). Summaries that fuse the two and add "a burst of dopamine" are stating a speculation as if it were the mechanism. Grade: good evidence for the effect; the dopamine story is speculative.

The difficulty has to be desirable: above your level, it is just difficulty. Vygotsky's zone of proximal development is the band between bored and frustrated, and the advice that follows from it is plain: when you are over the top edge, ask.

Po polsku, w skrócie¶

Wspomnienie przechodzi przez uwagę, pamięć roboczą, utrwalanie i przechowywanie, a każdy etap ma swoje ograniczenie: czego nie zauważyliśmy, tego nie zakodowaliśmy; pamięć robocza mieści około czterech porcji; świeży ślad łatwo nadpisać, a sen go utrwala; ślad blednie, jeśli się go nie używa. Cztery techniki o najmocniejszych dowodach działają każda na jedno z tych ograniczeń: sprawdzanie się z pamięci, powtórki rozłożone w czasie, przeplatanie i łączenie nowego ze znanym. Wszystkie są „pożądanymi trudnościami” (Bjork): w trakcie nauki idzie gorzej, na teście lepiej, dlatego uczniowie je niedoceniają. Krzywa zapominania Ebbinghausa mówi, że bez powtórki połowa znika w dzień lub dwa. Zgadywanie przed nauką pomaga nawet wtedy, gdy się mylimy; „hiperkorekcja” to co innego: błędy popełnione z dużą pewnością najłatwiej poprawić.