How Memory Works

In brief

Ask what memory is and the honest answer is that it is not one thing. To remember is to do three separable jobs: encode an experience into some internal code, store that code over time, and retrieve it when it is needed. Each job can succeed or fail independently, which is why a name can sit "on the tip of the tongue," fully stored yet momentarily unreachable. The two great frameworks of the field disagree about the architecture behind these jobs. Richard Atkinson and Richard Shiffrin (1968) proposed that information flows through a chain of distinct stores, sensory, short-term, and long-term, each with its own capacity and duration. Fergus Craik and Robert Lockhart (1972) countered that the stores are the wrong unit of analysis: what determines whether something is remembered is how deeply it is processed at encoding. Around both sits the older question of why memory fades, answered by four rival theories of forgetting.

The full treatment

The problem it answers

Before the 1950s, psychology had a rich record of how quickly memory decays, from Hermann Ebbinghaus onward, but no agreed picture of the system that produces it. The cognitive revolution reframed the mind as an information processor, and the pressing question became structural: is memory a single reservoir, or a set of stages with different properties? The evidence that forced the second answer was clinical. In 1957 William Scoville and Brenda Milner reported the patient H.M. (Henry Molaison), whose hippocampus and nearby tissue had been surgically removed to control epilepsy. Afterward he could hold a conversation and repeat a string of digits, his immediate memory intact, yet he could form almost no new lasting memories. Something kept short-term function while destroying the ability to make long-term traces. That dissociation demanded a model with separable parts.

How the multi-store model works

Atkinson and Shiffrin's answer, often called the modal model, is a conveyor belt. Sensory memory holds a near-complete but fleeting copy of what the senses just registered, visual traces for a fraction of a second, auditory ones for a few seconds. George Sperling's 1960 partial-report experiments showed that people briefly hold far more than they can report before it fades. Attention transfers a sliver of this into short-term memory, a small, fragile workspace. George Miller's 1956 paper had already fixed its capacity near "the magical number seven, plus or minus two" items, and Lloyd and Margaret Peterson showed in 1959 that without rehearsal its contents vanish within about eighteen seconds. Rehearsal both holds material in the short-term store and, the model claims, transfers it to long-term memory, a store of effectively unlimited capacity and duration. The serial position curve is the model's signature evidence: given a list, people recall the first items well (rehearsed into long-term memory) and the last items well (still in short-term memory), a pattern Glanzer and Cunitz dissociated in 1966 by showing a delay wipes out only the recency portion.

What levels of processing claims instead

Craik and Lockhart argued that the modal model put the emphasis in the wrong place. Rehearsal, they noted, does not reliably build durable memory: repeating a phone number over and over ("maintenance rehearsal") holds it briefly but leaves little trace an hour later. What predicts durable memory is the depth of processing at encoding. Analyze a word for its shape (shallow), its sound (intermediate), or its meaning (deep), and only deep, semantic processing produces strong recall. Craik and Tulving's 1975 experiments made the case: asking whether a word fit a sentence produced far better later memory than asking whether it was printed in capitals, even when the shallow task took longer. On this view there are not separate boxes so much as a continuum of encoding, and the trace is a byproduct of how thoroughly the mind worked on the input.

The theories of forgetting

Why do stored memories become unavailable? Four accounts compete, and they are not mutually exclusive. Decay theory holds that traces simply fade with time if unused, the intuition behind Ebbinghaus's forgetting curve, though pure time-based decay is hard to isolate from what happens during the interval. Interference theory locates the cause in competition between memories: retroactive interference is new learning disrupting old (a new phone number crowding out the old one), and proactive interference is old learning disrupting new. Retrieval-failure theory, associated with Endel Tulving, holds that the memory is intact but the cue needed to reach it is missing; his encoding specificity principle (with Donald Thomson, 1973) states that recall succeeds to the degree that the cues present at retrieval match those present at encoding, which is why returning to a place can flood back memories made there. Finally, motivated forgetting, in Sigmund Freud's version repression, holds that painful memories are actively kept from awareness. Once purely clinical and hard to test, this idea gained an experimental foothold in Michael Anderson and Collin Green's "think/no-think" studies from 2001, which found that deliberately suppressing retrieval of a learned item can make it harder to recall later.

Lineage

The scientific study of memory begins with Hermann Ebbinghaus (1850 to 1909), who in 1885 memorized lists of nonsense syllables on himself and plotted the first forgetting curve, showing that most loss happens fast and then levels off. He established that memory could be measured at all. Wilhelm Wundt's founding of experimental psychology supplied the laboratory method that made such work possible. The stage model itself has an older ancestor in William James (1842 to 1910), who in 1890 distinguished "primary memory," the just-past still held in consciousness, from "secondary memory," the stored past that must be retrieved. Atkinson and Shiffrin formalized James's distinction into an information-processing architecture, and Craik and Lockhart then pushed against the architecture in favor of process.

The strongest case for it

The multi-store model earns its long dominance by explaining a wide spread of findings with a simple structure. It predicts the serial position curve and its dissociation under delay. It accounts for H.M. and similar amnesic patients, whose intact short-term span alongside a devastated ability to form new long-term memories is exactly the double dissociation separate stores predict. And its central distinction, that immediate holding and durable storage are different systems, has survived essentially every later revision. The levels-of-processing framework, in turn, corrects the model's weakest claim, that rote rehearsal builds memory, and replaces it with a robust and often-replicated finding: meaning is what makes things stick. Taken together the two frameworks give teachers and students something usable. The reason spaced, effortful, meaningful practice beats cramming, documented in the entry on the spacing effect, follows directly from deep encoding plus distributed retrieval.

The strongest case against it

The modal model has been attacked on its own terms. Alan Baddeley and Graham Hitch argued in 1974 that the unitary short-term store cannot be right: it fails to explain how people simultaneously hold information and manipulate it, and they replaced it with a multi-component "working memory," a controlling executive plus separate stores for verbal and visual material. Neuropsychological cases sharpened the point. Patient K.F., studied by Tim Shallice and Elizabeth Warrington in 1970, had a badly impaired short-term span yet could still form new long-term memories, the reverse of what a strict serial model predicts if everything must pass through short-term memory to reach long-term storage.

Levels of processing drew its own strong objections. Its central term is circular: "depth" is defined by what produces good memory, and good memory is then explained by depth, so the theory risks predicting nothing it did not assume. Critics including Craik himself acknowledged the difficulty of measuring depth independently. Later work by Morris, Bransford, and Franks (1977) showed that deep semantic processing is not always best; under "transfer-appropriate processing," a shallow rhyme-based encoding produced better memory when the later test itself demanded rhyme judgments. What matters is the match between how you encode and how you will be tested, not depth as such, which reframes the finding rather than confirming it.

Motivated forgetting remains the most contested piece. The "recovered memory" controversy of the 1990s, and Elizabeth Loftus's work showing that entire false memories can be experimentally implanted, made many researchers wary of repression as an explanation for forgotten trauma, warning that a memory can feel vivid and recovered yet never have occurred.

Where it stands now

No single framework won. The field's working picture is a synthesis. The gross architecture, a distinction between transient immediate memory and durable long-term memory, is not seriously in doubt and is anchored in the biology of the hippocampus and its consolidation of memories over time. But the simple short-term box has been replaced by Baddeley's working memory, and long-term memory is now subdivided (Endel Tulving's episodic versus semantic; the explicit versus implicit split that H.M. revealed when he improved at motor tasks he could not remember practicing). Levels of processing survives not as a rival theory but as a durable principle folded into encoding research: meaning and appropriate matching of encoding to retrieval drive durability. On forgetting, the modern consensus is pluralist. Interference and retrieval failure carry most of the everyday load, time-based decay plays a smaller and disputed role, and active suppression is real but narrower than Freud supposed. Memory, the field now holds, is less a recording than a set of reconstructive processes, which is why it can be both indispensable and, at times, confidently wrong.

Test yourself

Think of something you studied hard for and have since lost, and something learned once that has never left you. Ask which stage failed in the first case: did you never encode it deeply, or is it stored but now uncued? Then ask what the second memory had that the first lacked, meaning, emotion, repeated retrieval over spaced intervals. The difference between the two is the whole of this entry, running in your own head.