Long-Term Memory

Abstract

Long-term memory is the brain's durable store, the system that retains knowledge, skills, and experience over intervals from minutes to a lifetime, in contrast to the fleeting hold of working memory. This article distinguishes its major systems, the declarative memory for facts and events and the nondeclarative memory for skills and conditioning, and follows a memory from encoding through consolidation to retrieval. It draws on the amnesic patient H.M. to locate the medial temporal lobe at the heart of declarative memory, traces consolidation from synaptic change to the slow reorganization of cortical networks, and examines why memory is reconstructive rather than a faithful recording. Three interactive demonstrations let the reader vary depth of encoding, watch a memory consolidate from hippocampus to cortex, and compare restudy against retrieval practice.

Keywords: long-term memory, consolidation, retrieval practice

Long-term memory is the system that holds information in a durable form after it has left the transient grip of attention and working memory, retaining facts, events, skills, and associations over intervals that can stretch across a lifetime (Atkinson & Shiffrin, 1968). It is defined less by any single time constant than by a contrast: where working memory holds a handful of items for seconds and loses them the moment attention moves on, long-term memory has no known ceiling on capacity and no fixed limit on duration (Cowan, 2001). The concept matters to cognitive psychology because it is not one faculty but several, dissociable in the brain and behaviorally distinct, and because the life of a memory, how it is encoded, made permanent, retrieved, and quietly rewritten in the process, has become one of the most productive research programs spanning psychology and neuroscience. The sections below distinguish the memory systems, locate declarative memory in the medial temporal lobe through the amnesic evidence, follow encoding and consolidation, and examine why remembering is an act of reconstruction rather than replay.

The Store Beyond the Moment

The place of long-term memory in the architecture of the mind was fixed by the modal model of Atkinson and Shiffrin, who divided memory into a sensory register, a limited short-term store, and an effectively unlimited long-term store, with information passing into the long-term store to the degree that it was rehearsed and processed in the short-term store (Atkinson & Shiffrin, 1968). The contrast with the short-term store is the defining feature. Immediate memory holds only a few items, famously around seven in George Miller's original estimate and closer to four once rehearsal and grouping are stripped away, and it loses them within seconds unless they are actively maintained (Miller, 1956; Cowan, 2001). Long-term memory has neither limit: no one has measured a ceiling on how much a person can know, and durable memories can survive decades without rehearsal.

The experimental study of that durability began with Hermann Ebbinghaus, who by learning and relearning lists of nonsense syllables charted the first forgetting curve, showing that retention falls steeply at first and then flattens (Ebbinghaus, 1913). At the far end of the scale, Harry Bahrick tracked memory for Spanish learned in school across fifty years and found that knowledge surviving the first few years then held almost unchanged for decades, a durable residue he named permastore (Bahrick, 1984). The transfer between the two is not automatic. Whether an experience leaves a lasting trace depends heavily on how it is processed at the moment of encoding, which is where the study of long-term memory properly begins, and it depends further on a period of stabilization after encoding during which the fragile new trace is at risk. The working memory that feeds the long-term store is treated in detail in the companion article on working memory. Table 1 sets the two stores side by side across the dimensions on which they differ.

Note. The two stores are distinguished less by any single property than by their convergence across all of them. Capacity and duration estimates follow the immediate-span tradition and its reconsideration (Miller, 1956; Cowan, 2001); the transfer between stores is the core of the modal model (Atkinson & Shiffrin, 1968).

Types of Long-Term Memory

Beyond being a subject in its own right, Long-Term Memory is a formal category in the National Library of Medicine's Medical Subject Headings, which places it at tree position F02.463.425.540.305, beneath Memory, and hangs its recognised narrower kinds beneath it. These subtypes are a classification built to index the literature, not a claim about the mind's natural joints; several are treated at length in their own articles. Table 2 lists the direct children of the descriptor.

Two cautions keep this taxonomy in its place. It is a classification for indexing, built to organise the literature, not a theory asserting these subtypes are mutually exclusive or an exhaustive set of natural kinds. And a MeSH subtype is a narrower topic, not a component process: listing a child under Long-Term Memory locates it in an index and says nothing, on its own, about the mechanisms this article describes.

Systems of Long-Term Memory

The single most important discovery about long-term memory is that it is not one thing. The major division is between declarative memory, the memory for facts and events that can be consciously brought to mind and stated, and nondeclarative memory, a collection of abilities expressed through performance rather than recollection, including motor and cognitive skills, priming, and simple conditioning (Squire, 2004). Within declarative memory, Endel Tulving drew a further distinction that has organized the field ever since: episodic memory, the recollection of specific personally experienced events set in a particular time and place, and semantic memory, the store of general knowledge about the world stripped of the context in which it was learned (Tulving, 2002). Remembering a specific breakfast is episodic; knowing that eggs are food is semantic. Tulving argued that episodic memory is distinctive in involving a kind of mental time travel, a re-experiencing of the past accompanied by autonoetic awareness that the event was one's own. Figure 1 lays out the taxonomy. These systems are not merely conceptual conveniences: they depend on partly different brain structures, which is why brain damage can devastate one while sparing another, the finding that first made the distinctions unavoidable.

Encoding: Getting In

Whether an experience becomes a lasting memory is decided largely at encoding, and the decisive variable is not effort or repetition but the depth at which the material is processed. Fergus Craik and Robert Lockhart proposed the levels-of-processing framework, arguing that memory is a byproduct of perception and comprehension: material analyzed shallowly for its surface features, such as the shape of a printed word, leaves a fragile trace, whereas material analyzed deeply for its meaning leaves a durable one, regardless of any intention to remember (Craik & Lockhart, 1972). Judging whether a word is printed in capitals produces poor later memory; judging whether it fits sensibly into a sentence produces far better memory, even though both take similar time. This reframed encoding as a matter of the qualitative operations performed on information rather than the sheer duration it is held. What is encoded, moreover, is not the event alone but the context in which it was processed. Tulving and Thomson's encoding-specificity principle holds that retrieval succeeds to the extent that the cues available at recall match those present at encoding, so that a memory is most accessible in the conditions under which it was formed (Tulving & Thomson, 1973). The two ideas together explain why elaborative, meaningful encoding that ties new material into a rich web of retrieval routes produces memory that both lasts and can be found again. The demonstration below varies the depth of encoding and shows its effect on later recall.

The Amnesic Evidence

The clearest window onto the organization of long-term memory came from a single patient. In an attempt to control his epilepsy, the surgeon William Scoville removed the medial temporal lobes, including most of the hippocampus, on both sides of the brain of a young man known for decades only as H.M. The seizures eased, but at a catastrophic cost that Scoville and Brenda Milner documented: a profound anterograde amnesia, an inability to form new lasting declarative memories, while his intelligence and personality were preserved and his older memories were largely spared, apart from a partial, temporally graded loss of the years just before the operation (Scoville & Milner, 1957). H.M. could hold a conversation but would forget it minutes later; he met his doctors as strangers for the rest of his life. The case established that the medial temporal lobe is not the site where memories are permanently stored but is necessary for forming new declarative memories, and that this function is separable from immediate memory and from intelligence. Decades of careful study, reviewed by Suzanne Corkin, refined the picture: H.M.'s deficit was specific to declarative memory, sparing his ability to acquire new motor skills, which he improved on across days even while denying he had ever attempted the task (Corkin, 2002). That dissociation, preserved skill learning alongside abolished fact-and-event learning, is the behavioral proof that declarative and nondeclarative memory are distinct systems. Larry Squire synthesized the human, monkey, and rodent evidence into an account of the medial temporal lobe memory system in which the hippocampus and adjacent cortices bind the elements of an experience into a coherent declarative memory and support it until it can become independent of them (Squire, 1992). The wider consequences of such damage are treated in the article on amnesia.

Consolidation: Making It Last

A memory is not permanent the moment it forms. It passes through a period of consolidation during which it is gradually stabilized, and during which it remains vulnerable to disruption, an idea James McGaugh traced across a century of research showing that treatments applied after learning, from drugs to emotional arousal, can strengthen or impair a memory precisely because the trace is still settling (McGaugh, 2000). Consolidation operates on two timescales. At the synaptic level, the cellular basis was illuminated by Timothy Bliss and Terje Lømo's discovery of long-term potentiation, a lasting increase in synaptic strength following brief high-frequency stimulation, which provided a plausible physical mechanism for how experience durably alters neural connections (Bliss & Lømo, 1973). At the systems level, consolidation is far slower. Paul Frankland and Bruno Bontempi described how declarative memories, initially dependent on the hippocampus, are gradually reorganized over weeks to years so that retrieval comes to rely increasingly on distributed neocortical networks, which explains why medial temporal damage spares remote memories while abolishing recent ones (Frankland & Bontempi, 2005). This standard account is contested, however. Lynn Nadel and Morris Moscovitch's multiple-trace theory holds that richly detailed episodic memories remain permanently dependent on the hippocampus, each retrieval laying down a fresh trace, so that the graded sparing of remote memory applies more to semanticized knowledge than to vivid recollection (Nadel & Moscovitch, 1997).

Much of this reorganization appears to take place during sleep. The hippocampus replays patterns of recent experience during slow-wave sleep, and this reactivation is thought to drive the gradual redistribution of memories to neocortical networks, which is why a night of sleep after learning reliably improves later retention (Diekelmann & Born, 2010). Consolidation also turned out to be less final than once believed. Karim Nader and colleagues showed that recalling a consolidated memory can return it to a labile state requiring protein synthesis to restabilize, a process termed reconsolidation, implying that each act of retrieval is an opportunity for a memory to be updated or altered (Nader, Schafe, & Le Doux, 2000). The demonstration below traces how a memory's dependence shifts from hippocampus to cortex over time.

Reconstruction and Distortion

Retrieval is not the replay of a stored recording but the reconstruction of an event from partial traces, general knowledge, and the cues present at the moment of remembering, and this reconstructive character makes memory systematically vulnerable to distortion. The idea that remembering is reconstructive rather than reproductive traces to Frederic Bartlett, whose readers recalling the unfamiliar folk tale The War of the Ghosts progressively reshaped it across successive retellings, dropping and rationalizing unfamiliar details to fit the schemas of their own culture (Bartlett, 1932). The most consequential experimental demonstration came from Elizabeth Loftus and John Palmer, who showed witnesses a film of a car accident and then varied a single word in the question they asked: those asked how fast the cars were going when they smashed into each other estimated higher speeds than those asked when the cars hit, and a week later were more likely to report having seen broken glass that was never in the film (Loftus & Palmer, 1974). The wording of the question had entered and altered the memory itself. This misinformation effect, together with the finding that retrieving a memory can return it to a modifiable state (Nader, Schafe, & Le Doux, 2000), reframes remembering as an active and error-prone process rather than a passive lookup. The practical stakes are high, most obviously for the reliability of eyewitness testimony, and the theoretical lesson is that the malleability of memory is not a defect of an otherwise perfect system but an intrinsic feature of a store built for meaning and generalization rather than verbatim fidelity.

Making Memories Durable

The reconstructive, consolidating character of long-term memory also points to how learning can be made to last, and here the laboratory findings are unusually clear and practical. Two robust effects stand out. The spacing effect holds that information studied in sessions distributed over time is retained far better than the same information crammed into a single massed session; Nicholas Cepeda and colleagues, synthesizing more than a century of studies, found not only that distributed practice reliably wins but that the optimal gap between sessions grows with the interval over which the material must be retained (Cepeda, Pashler, Vul, Wixted, & Rohrer, 2006). The testing effect, or retrieval practice, holds that the act of retrieving a memory strengthens it more than an equivalent amount of restudy; Henry Roediger and Jeffrey Karpicke found that although rereading produces better performance on an immediate test, taking a practice test produces substantially better retention when the delay is a week or more, precisely the interval that matters for durable learning (Roediger & Karpicke, 2006). Both effects follow from the same principles that structure the rest of this article: retrieval is reconstructive, so exercising it builds the routes that make a memory findable, and consolidation is slow, so spacing gives each trace time to stabilize before it is strengthened again. Deliberate mnemonic techniques such as the method of loci exploit the same logic of elaborate, well-cued encoding. The demonstration below compares restudy against retrieval practice as the retention interval grows.

Worked Example

The reversal at the heart of the testing effect can be made quantitative with the illustrative forgetting model in the demonstration above, in which the proportion recalled after t days decays exponentially: restudy follows R equals 0.81 times e raised to the power of negative t divided by 10.7, and retrieval practice follows R equals 0.75 times e raised to the power of negative t divided by 24.0. Retrieval practice starts lower, 0.75 against 0.81, because a single test session leaves less immediately accessible than rereading does, but it forgets more slowly, because the longer time constant of 24.0 days reflects the durability that retrieval builds. Evaluate both at one week. Restudy gives 0.81 multiplied by e to the power of negative 7 divided by 10.7, which is 0.81 multiplied by 0.520, or about 0.42. Retrieval practice gives 0.75 multiplied by e to the power of negative 7 divided by 24.0, which is 0.75 multiplied by 0.747, or about 0.56. The ranking has reversed: the method that looked worse immediately now retains half again as much. The crossover point, where the two are equal, is found by setting the curves equal and solving for t, which gives t equals the natural logarithm of 0.81 divided by 0.75, all divided by the quantity one over 10.7 minus one over 24.0. That is 0.0770 divided by 0.0518, or about 1.5 days. In words, retrieval practice overtakes restudy after roughly a day and a half, and by one week its advantage is a net 0.56 minus 0.42, or 0.14, a relative gain of about 33 percent. The numbers are illustrative, but the shape, an early cost repaid many times over at the delays that matter, is exactly what the empirical testing effect shows.

Discussion

Long-term memory matters first because it is the foundation of nearly everything else in cognition: knowledge, expertise, personal identity, and the ability to plan for a future all rest on a store that outlasts the moment. It matters second because the discovery that this store is fractionated, declarative against nondeclarative, episodic against semantic, dissociable structure by structure in the brain, is one of the genuine triumphs of cognitive psychology and neuroscience, built on the convergence of the amnesic evidence with animal models and, later, neuroimaging (Squire, 2004). It matters third because the life cycle of a memory, encoded by depth of processing, stabilized by consolidation over hours and years, and reconstructed rather than replayed at retrieval, has turned out to be a rich and unified story rather than a set of disconnected facts, one in which the same reconstructive nature that makes memory fallible also makes it improvable through spacing and retrieval practice (Cepeda, Pashler, Vul, Wixted, & Rohrer, 2006; Roediger & Karpicke, 2006). The open questions remain deep. How a distributed cortical memory is indexed and bound, whether reconsolidation genuinely rewrites a trace or merely adds to it, and how the durable permanence of remote memory is achieved at the molecular level are all active fronts. What is settled is that remembering is a constructive achievement of a living system, not the readout of a recording, and that this single fact organizes most of what the field has learned.

Common Misconceptions

Memory works like a video recording that can be played back.

Retrieval reconstructs an event from partial traces and current cues, and the reconstruction can absorb information encountered after the fact. Loftus and Palmer showed that merely changing a verb in a question raised witnesses' speed estimates and later led them to remember broken glass that was never present (Loftus & Palmer, 1974). A recording does not rewrite itself when questioned; memory does.

Long-term memories are permanent as soon as they form.

A new trace is fragile and stabilizes only gradually through consolidation, remaining open to disruption for hours at the synapse and reorganizing across neocortex over years (McGaugh, 2000; Frankland & Bontempi, 2005). Recall can even return a consolidated memory to a labile state (Nader, Schafe, & Le Doux, 2000). Permanence is achieved slowly, not conferred at encoding.

Rereading is the most effective way to make material stick.

Rereading feels effective and does help on an immediate test, but at the delays that matter for durable learning, actively retrieving the material outperforms restudying it (Roediger & Karpicke, 2006). Distributing that practice over time beats massing it into one session (Cepeda, Pashler, Vul, Wixted, & Rohrer, 2006). The comfortable strategy is not the effective one.

Glossary

Amnesia.

A loss of memory from brain injury or disease; anterograde amnesia impairs forming new memories, retrograde amnesia impairs recalling old ones.

Consolidation.

The process by which a new, fragile memory is gradually stabilized, at the synapse over hours and across cortical networks over years.

Declarative memory.

Memory for facts and events that can be consciously brought to mind and stated; also called explicit memory.

Encoding specificity.

The principle that retrieval succeeds to the degree that the cues available at recall match those present when the memory was encoded.

Episodic memory.

The recollection of specific personally experienced events set in a particular time and place, accompanied by a sense of reliving.

Hippocampus.

A medial temporal lobe structure essential for forming new declarative memories and for binding the elements of an experience.

Levels of processing.

The framework holding that memory strength depends on the depth of processing at encoding, with meaning-based analysis leaving the most durable trace.

Long-term memory.

The durable memory store that retains knowledge, events, and skills over intervals from minutes to a lifetime, with no fixed limit on capacity or duration.

Long-term potentiation.

A lasting increase in synaptic strength following brief high-frequency stimulation, a leading cellular mechanism for durable memory.

Nondeclarative memory.

A collection of memory abilities expressed through performance rather than conscious recollection, including skills, priming, and conditioning; also called implicit memory.

Reconsolidation.

The return of a reactivated memory to a labile state that requires restabilization, allowing an established memory to be updated or altered.

Reconstruction.

The building of a remembered event from partial traces, general knowledge, and current cues, rather than the replay of a stored recording.

Semantic memory.

The store of general knowledge and facts about the world, abstracted away from the specific context in which it was learned.

Spacing effect.

The finding that information studied in sessions distributed over time is retained better than the same study massed into one session.

Testing effect.

The finding that retrieving a memory through testing strengthens later retention more than an equivalent amount of restudy; also called retrieval practice.

Key Researchers

Endel Tulving (1927–2023). Estonian-born Canadian psychologist at the University of Toronto and the Rotman Research Institute; he introduced the distinction between episodic and semantic memory and the encoding-specificity principle, shaping the modern understanding of how memories are organized and retrieved.
Google Scholar · University of Toronto · Wikipedia

Brenda Milner. Professor of neuroscience at the Montreal Neurological Institute and McGill University; her study of the patient H.M. established that the medial temporal lobe is necessary for forming new declarative memories and that skill learning is spared, founding the cognitive neuroscience of memory.
Faculty Page · Wikipedia

Larry R. Squire. Distinguished Professor at the University of California, San Diego and the VA San Diego Healthcare System; he synthesized the human and animal evidence into the account of the medial temporal lobe memory system and the declarative-nondeclarative taxonomy.
Faculty Page · ORCID · Google Scholar · Wikipedia

Fergus I. M. Craik. Senior scientist emeritus at the Rotman Research Institute, Baycrest, and the University of Toronto; with Robert Lockhart he proposed the levels-of-processing framework, recasting memory as a byproduct of the depth at which information is processed.
Faculty Page · ORCID · Google Scholar · Wikipedia

James L. McGaugh. Research professor at the University of California, Irvine and founder of its Center for the Neurobiology of Learning and Memory; his work established that memory consolidation can be modulated after learning by arousal and neurochemical systems.
Faculty Page · Google Scholar · Wikipedia

Elizabeth F. Loftus. Distinguished Professor at the University of California, Irvine; her experiments on the misinformation effect demonstrated the reconstructive, suggestible nature of memory and transformed the understanding of eyewitness reliability.
Faculty Page · ORCID · Google Scholar · Wikipedia

Frequently Asked Questions

What is long-term memory?
Long-term memory is the durable store that retains knowledge, events, and skills after they have left working memory, over intervals ranging from minutes to a lifetime. Unlike the few items held for seconds in short-term memory, it has no measured ceiling on capacity and no fixed limit on how long a memory can last (Atkinson & Shiffrin, 1968).

How is long-term memory different from short-term memory?
Short-term or working memory holds only a handful of items, roughly four once rehearsal is controlled, and loses them within seconds unless they are actively maintained. Long-term memory holds vast amounts of information durably and does not require continuous attention to keep it, which is why the two are treated as distinct systems (Cowan, 2001).

What are the main types of long-term memory?
The major division is declarative memory, for consciously accessible facts and events, and nondeclarative memory, for skills, priming, and conditioning expressed through performance. Declarative memory divides further into episodic memory for personal events and semantic memory for general knowledge (Squire, 2004; Tulving, 2002).

What did the patient H.M. teach us about memory?
After surgical removal of his medial temporal lobes, H.M. could form no new declarative memories yet learned new motor skills normally and kept his older memories and intelligence. This proved that the medial temporal lobe is necessary for forming new facts and events, and that declarative and skill memory are separate systems (Scoville & Milner, 1957; Corkin, 2002).

What is memory consolidation?
Consolidation is the process by which a new, fragile memory becomes stable and lasting. It works at two timescales: rapid changes in synaptic strength over hours, and a slow reorganization over months to years in which memories become less dependent on the hippocampus and more reliant on cortical networks (McGaugh, 2000; Frankland & Bontempi, 2005).

Why is memory unreliable?
Because retrieval reconstructs an event from partial traces and current cues rather than replaying a recording, memory can absorb information encountered after the event. Loftus and Palmer showed that the wording of a question could raise witnesses' speed estimates and implant a detail that was never seen (Loftus & Palmer, 1974).

What is the best way to make learning last?
Two strategies are strongly supported: distributing study across spaced sessions rather than massing it, and practicing retrieval by testing rather than only rereading. Both outperform the more comfortable alternatives at the delays that matter for durable learning (Cepeda, Pashler, Vul, Wixted, & Rohrer, 2006; Roediger & Karpicke, 2006).

Does the brain have a place where memories are stored?
There is no single storehouse. The hippocampus is needed to form new declarative memories but is not their permanent home; over time memories come to depend on distributed neocortical networks, and different systems such as the basal ganglia and cerebellum support skills and conditioning (Squire, 1992).

References