Learning and Memory Systems

Memory researchers distinguish two broad superordinate categories, first formalized clearly in these terms by Larry Squire in the 1980s building on the neuropsychology of H.M.: declarative memory and nondeclarative memory. Declarative memory (also called explicit memory) is memory for facts and events that can be brought to conscious awareness and, crucially, can be declared — stated in a sentence, compared, and reasoned about. It further splits into episodic memory (memory for specific personal events situated in time and place — what you had for dinner last night) and semantic memory (general knowledge and facts stripped of the episode in which they were learned — the capital of France). Declarative memory is fast: a single trial, or even a fraction of a second's exposure, can create a lasting declarative memory (you can meet someone once and remember their name years later).

Nondeclarative memory (also called implicit memory) is an umbrella term for several unrelated abilities that improve with experience but do not require, and are not accessible to, conscious recollection. It includes procedural memory (motor and cognitive skills, such as riding a bicycle or reading mirror-reversed text), priming (facilitated processing of a stimulus due to prior exposure, without awareness of that exposure), classical/Pavlovian conditioning (learned associations between stimuli, including emotional conditioning and reflex conditioning), and nonassociative learning (habituation and sensitization, the simplest forms of learning, extensively characterized in Aplysia by Eric Kandel's group). Nondeclarative memory is typically slow and incremental, built up over many trials, and it is expressed only through performance — you cannot "tell" someone your improved golf swing, you can only demonstrate it. The theoretical payoff of this taxonomy is that it is not just a descriptive convenience — it maps onto genuinely separate neural hardware. Declarative memory formation depends critically on the hippocampus and adjacent medial temporal lobe cortex, structures that bind together the distributed neocortical representations of an experience into a single retrievable memory trace, then gradually transfer stable versions of that memory to neocortex over a consolidation period. Nondeclarative memory forms through entirely separate circuits — the striatum for habit learning, the cerebellum for motor learning and delay eyeblink conditioning, and the amygdala for fear conditioning — that do not require the hippocampus at all. This is why a single lesion (H.M.'s) could destroy one whole category of learning while leaving another intact: the underlying anatomy really is separate, not just conceptually distinguishable. Understanding this separation depends on prior work establishing how synapses physically consolidate — see Molecular Mechanisms of Memory Consolidation — but the systems-level story adds the crucial claim that where consolidation happens differs by memory type.

Step 1 — Encoding in the medial temporal lobe (declarative pathway). Sensory and associative information about an ongoing experience arrives at the hippocampus from widely distributed regions of neocortex (visual, auditory, spatial, emotional), funneled through three adjacent cortical relay stations: the perirhinal cortex (object and item information), the parahippocampal cortex (spatial/contextual information), and finally the entorhinal cortex, which is the main gateway into and out of the hippocampus proper. Within the hippocampus, the classic trisynaptic circuit — dentate gyrus → CA3 → CA1 — performs pattern separation (making similar experiences distinguishable) and pattern completion (retrieving a whole memory from a partial cue), using NMDA-receptor-dependent long-term potentiation at these synapses as the basic plasticity mechanism.

Step 2 — Binding. The hippocampus does not store the sensory content of the memory itself. Instead it rapidly binds together the separate cortical modules that represent a single experience (a face, a location, a smell, a time) into one index or pointer. This is why hippocampal damage does not erase old skills or old semantic facts, but does prevent new experiences from being unified into retrievable episodes.

Step 3 — Systems consolidation. Over subsequent hours to years, that hippocampal index is replayed — notably during slow-wave sleep, in coordinated hippocampal sharp-wave ripples synchronized with neocortical slow oscillations and thalamic spindles — gradually strengthening direct neocortical-neocortical connections until the memory can be retrieved without the hippocampus at all. This explains temporally graded retrograde amnesia: hippocampal damage disrupts recent memories (still hippocampus-dependent) far more than remote memories (already consolidated into neocortex), a pattern documented in H.M. and formalized as the standard model of systems consolidation by Larry Squire and colleagues.

Step 4 — Nondeclarative pathways bypass the hippocampus entirely. In parallel, and using entirely different circuitry:

• Procedural/habit learning depends on cortico-striatal loops: sensory and motor cortex project to the striatum (caudate, putamen), which uses dopamine-modulated synaptic plasticity to gradually strengthen stimulus–response associations through repetition, independent of the hippocampus.
• Motor skill learning and delay eyeblink conditioning depend on the cerebellum, where climbing fiber input from the inferior olive provides an error/teaching signal that adjusts Purkinje cell synaptic weights (long-term depression at parallel fiber–Purkinje synapses), a mechanism directly demonstrated by Richard Thompson's lesion and recording studies in rabbits.
• Fear/emotional conditioning depends on the amygdala, where convergence of conditioned stimulus (tone) and unconditioned stimulus (shock) pathways onto lateral amygdala neurons produces associative LTP, established by Joseph LeDoux's and Michael Davis's work.
• Priming depends on changes within neocortical perceptual and conceptual processing areas themselves (reduced neural firing for repeated stimuli, "repetition suppression"), requiring no subcortical structure and no hippocampus.

Step 5 — The double dissociation. Because these are anatomically separate systems, damage to one need not affect the other. Damage to the medial temporal lobe (H.M.) destroys declarative memory formation but spares procedural, conditioning, and priming. Damage to the striatum (Huntington's or Parkinson's disease) impairs habit learning but spares declarative memory. Damage to the cerebellum impairs motor-skill and eyeblink conditioning but spares both declarative memory and habit learning. Damage to the amygdala selectively impairs the emotional coloring/fear conditioning component of a memory while sparing the factual/declarative record of the same event. This last dissociation was demonstrated directly by Antonio Damasio, Ralph Adolphs, and colleagues in a patient (S.M.) with bilateral amygdala damage (Urbach–Wiethe disease) who could describe the facts of a conditioning session perfectly but showed no autonomic fear response, versus a hippocampal patient who showed the fear response but could not describe the session.

The case of H.M. (Henry Molaison), studied by Brenda Milner and later William Scoville, 1957 onward. Scoville, the neurosurgeon, and Milner, the neuropsychologist, published the foundational 1957 paper describing nine patients who had undergone medial temporal lobe resections for psychiatric or epileptic indications; H.M.'s case was by far the most striking and became the most intensively studied single patient in the history of neuroscience. His resection removed the hippocampus, amygdala, and adjacent entorhinal and perirhinal cortex bilaterally (roughly the anterior two-thirds of the hippocampus on each side, according to later MRI reconstructions by Suzanne Corkin's group). What was observed: H.M. developed a profound and permanent anterograde amnesia — a total inability to form new declarative memories — while retaining:

• Normal short-term/working memory (he could hold a number in mind and repeat it back, showing a normal digit span around 5–7 items, as long as rehearsal was uninterrupted).
• Normal intelligence (IQ actually rose slightly after surgery, since seizures had stopped) and normal language.
• Remote memories from well before the surgery, especially childhood, largely intact, though memories from roughly the 1–2 years immediately preceding surgery were also impaired (temporally graded retrograde amnesia).
• Fully intact ability to learn new motor skills. In the signature experiment, Milner had H.M. trace the outline of a five-pointed star while watching his hand only in a mirror — a task that is initially very difficult because it inverts visual feedback. Across three days of trials, H.M.'s error rate and time-to-completion improved steadily, at a rate indistinguishable from healthy control subjects — even though, at the start of each daily session, he denied ever having done the task before.
• Preserved priming: shown a fragmented picture or a word stem, H.M. was more likely to complete it correctly if he had seen the whole item earlier, even though he had no conscious memory of the earlier exposure.
• Preserved simple classical conditioning in later studies (e.g., conditioned eyeblink responses could still be acquired, albeit sometimes more slowly than normal).

What this proved: memory is not unitary. A single, well-localized lesion abolished one entire category of learning while sparing others completely, which is only possible if those categories depend on separate neural machinery. This was the first clear demonstration that the hippocampal/medial temporal lobe system is necessary specifically for the conscious, declarative, one-trial, context-bound form of memory, and is not required at all for skill learning, priming, or basic conditioning.

The complementary double dissociation — basal ganglia disease. If the MTL/striatum split were a matter of degree rather than kind, one might expect all forms of memory to be somewhat impaired together whenever any memory system is damaged. This possibility was ruled out by studies of patients with Huntington's disease (which principally damages the caudate and putamen, sparing the hippocampus) and Parkinson's disease. These patients show the opposite profile from H.M.: they are impaired at gradually-learned habit and skill tasks (such as the probabilistic classification "weather prediction" task developed by Barbara Knowlton, Squire, and colleagues, and at rotary-pursuit motor skill learning), yet their declarative memory for facts and single events can be relatively preserved, especially early in the disease. Neal Cohen and Squire's comparative studies through the 1980s directly contrasted amnesic patients (MTL damage) with Huntington's patients on the same tasks and showed this crossover pattern.

Amygdala versus hippocampus — the S.M. study (Bechara, Tranel, Damasio, Adolphs, Rockland, Damasio, 1995, Science). Three patients were compared: S.M., who had bilateral amygdala damage from Urbach–Wiethe disease but an intact hippocampus; W.C., who had bilateral hippocampal damage but an intact amygdala; and healthy controls. All underwent a simple fear conditioning procedure pairing a colored slide with a loud boat horn, while skin conductance response (a physiological fear measure) was recorded. Controls and W.C. (hippocampal damage) both developed a conditioned skin-conductance response to the colored slide, and could later declare the fact that the blue slide had been paired with the noise — except W.C. could not, since his declarative memory for the pairing itself was impaired despite his intact physiological conditioning. S.M. showed the mirror-image pattern: she could declare correctly, in detail, that the blue slide had predicted the horn, but she never developed a measurable conditioned skin-conductance response at all. This double dissociation demonstrated cleanly that the amygdala is necessary for acquiring the emotional/autonomic component of a conditioned memory, while the hippocampus is necessary for the declarative, verbalizable record of the same learning episode — two parallel traces of one event, laid down by two different structures.

Example 1 — Localizing a lesion from a memory profile. A 68-year-old patient presents after a stroke affecting the medial temporal lobe. Neuropsychological testing shows:

• Cannot recall a list of 15 words after a 10-minute delay (score: 1/15, controls average ~9/15).
• Cannot recall the name of a nurse introduced 20 minutes earlier.
• Learns the 10-item Tower of Hanoi puzzle across five days, halving her move count from ~60 to ~31, matching healthy age-matched controls (~65 to ~28).
• Shows normal word-stem completion priming (58% primed completions vs. 30% baseline guessing, matching controls' 55–60%).
• Digit span is normal (forward span 6, backward span 4).

Reasoning: The intact digit span rules out a basic attention/working-memory deficit — she can hold information briefly. The severe 10-minute free recall failure, contrasted with fully preserved procedural learning (Tower of Hanoi) and preserved priming, is the classic H.M. profile: an isolated declarative memory deficit. This localizes the lesion to hippocampal/medial temporal lobe circuitry, sparing striatal and neocortical priming systems, without needing an MRI to guess the location — the dissociation pattern itself is diagnostic, exactly the logic used clinically since the 1970s.

Example 2 — Quantifying consolidation gradient (a retrograde amnesia calculation). A researcher tests a hippocampal-lesion patient's memory for public events from five time periods before injury: 1–2 years, 3–5 years, 6–10 years, 11–15 years, and >20 years pre-injury. Recall accuracy (percent of events correctly recognized, controls normalized to 100%) is:

Reasoning: Control accuracy is roughly flat (~83–88%) across all time periods, meaning the events themselves are equally memorable regardless of age — a reasonable baseline. The patient's accuracy, in contrast, rises steadily from 22% (most recent) to 91% (most remote), approaching control levels only for memories more than ~15–20 years old. This is a textbook temporally graded retrograde amnesia: recent memories, which still depend on an intact hippocampus for retrieval, are devastated; remote memories, already fully transferred to neocortical storage via systems consolidation, are nearly spared. The gradient itself is quantitative evidence for a time-limited role of the hippocampus in each individual memory, exactly the pattern Squire's group documented in H.M. and in later hippocampal-lesion patients, and it directly supports the systems-consolidation model described in the mechanism section above.

Brenda Milner (born 1918) is the neuropsychologist most responsible for turning H.M.'s tragedy into rigorous science. Trained in experimental psychology at Cambridge under Oliver Zangwill, she moved to Montreal to work with neurosurgeon Wilder Penfield at the Montreal Neurological Institute, where she developed exacting behavioral methods for probing memory in neurosurgical patients. When William Scoville, the Hartford surgeon who had operated on H.M., recognized the severity of the resulting amnesia, he invited Milner to study the patient systematically. Beginning in 1955 and continuing for decades (H.M. traveled to Montreal repeatedly, and later researchers, including Suzanne Corkin, worked with him at MIT until his death in 2008), Milner's patient testing established, trial by careful trial, exactly what H.M. could and could not learn — including the pivotal 1962 mirror-tracing demonstration that first revealed spared procedural learning inside a profoundly amnesic brain. Her insistence on precise, quantified, repeated behavioral testing — rather than anecdote — turned a single tragic case into the empirical foundation for the entire declarative/nondeclarative taxonomy used across neuroscience today. She continued active research past age 100.

The declarative/nondeclarative framework is not just historical trivia — it actively shapes diagnosis and treatment. In Alzheimer's disease, pathology begins in entorhinal cortex and hippocampus, producing an H.M.-like profile of severe new declarative memory loss with relatively preserved procedural learning early on; this is why early Alzheimer's patients can still perform familiar routines and learn some new motor habits even as they cannot remember a conversation from an hour before, a distinction used in both diagnosis and in designing daily-living supports that lean on spared procedural and habit systems. In Parkinson's and Huntington's disease, damage to striatal circuits produces the mirror pattern, impaired habit/skill learning with relatively preserved declarative memory, guiding rehabilitation strategies (for instance, favoring explicit verbal instruction over trial-and-error practice for motor rehabilitation in Parkinson's patients, since their implicit learning system is compromised). In addiction neuroscience, the same striatal/habit system that supports normal skill learning is understood to be hijacked by drugs of abuse, explaining why habitual drug-seeking can become resistant to conscious, declarative knowledge that the behavior is harmful — a systems-level account, not a moral failing. In PTSD and anxiety disorder treatment, the amygdala-dependent, hippocampus-independent nature of fear conditioning explains why purely verbal/cognitive therapy can leave physiological fear responses intact, motivating exposure-based and extinction-based therapies that directly engage the conditioning circuitry rather than only the declarative narrative. Finally, in neurosurgical planning for epilepsy today, lessons from H.M.'s case directly drive practice: surgeons now use unilateral resections, intraoperative memory (Wada) testing, and detailed presurgical mapping specifically to avoid ever reproducing his outcome.

Ask the tutor: How does the standard model of systems consolidation explain why some remote memories eventually become "hippocampus-independent," and is this transfer ever complete?