Memory Systems

In 1953 a surgeon removed tissue from the inner part of both temporal lobes of a young man with severe epilepsy. Afterwards he could not remember events from an hour earlier, yet he could still learn a new hand skill.

That patient, H.M., is the reason you can ask a question that once seemed odd: is memory one ability or several? This lesson works through the evidence that it is several, and through the limits of that evidence, because nearly all of it comes from damaged brains.

The idea: one operation and a distinction

Start with the categories. Declarative memory is memory you can bring to conscious recollection and state. It comes in two kinds. Episodic memory is memory for events placed in a time and a place: what happened at breakfast. Semantic memory is general knowledge of the world: that a spider has eight legs. Nondeclarative memory is expressed in performance rather than recollection. It includes skills and habits, priming (being faster or more accurate with something you met recently), and simple classical conditioning. Working memory is different again: the temporary, active holding and manipulating of a few items, such as a phone number you repeat while you dial it.

An analogy: declarative memory is a diary you can read aloud, nondeclarative memory is the calluses and reflexes a job leaves on your hands, and working memory is a notepad you wipe every few seconds. The analogy is only useful if it can be tested, and the test is a dissociation: one ability fails after damage while another survives. Nondeclarative does not mean motor. As you will see, some of it is about predicting outcomes and some is about the heart rate and sweat glands.

The structure at the centre of this lesson is the medial temporal lobe, the inner (medial) part of each temporal lobe, which holds the hippocampus, the amygdala and the surrounding entorhinal, perirhinal and parahippocampal cortex. The nervous system map places it. Bilateral means on both sides.

What H.M. kept

William Scoville and Brenda Milner described H.M. and other patients from Scoville’s operations in 1957 in the Journal of Neurology, Neurosurgery and Psychiatry. H.M.’s main loss was anterograde amnesia: he could not form lasting new declarative memories. He also lost some memories from before the operation, a retrograde amnesia, though how far back it reached has been estimated differently over the years and is not given here.

The list of what he kept matters as much. His language and perception were retained. Formal intelligence testing did not show a general decline. His digit span, the number of digits he could repeat back immediately, was normal, and he could hold information as long as he kept rehearsing it. That separates working memory from long-term declarative memory: the notepad worked, but nothing written on it reached the diary. He learned new motor skills. In Milner’s later testing he improved at mirror drawing across three days while not remembering having practised. Over many years he also acquired, slowly, a few new facts. So “he could learn nothing new” is wrong twice over.

What H.M. shows is that some part of the medial temporal lobe is needed to form new declarative memories. It does not show that those memories are stored in the hippocampus. His factual knowledge of early life was largely spared, though he lost some memories from the years before the operation and later testing found his recollection of personal episodes poor across all remote periods, which is one reason the hippocampus is described as needed to form declarative memories and, for a time, to retrieve them. How that dependence changes with time is the subject of consolidation.

What the 1957 lesions could not isolate

The 1957 paper compared operations of different extent, and a lasting loss of recent memory went with resections that took in the anterior hippocampus and the hippocampal gyrus. But every resection that reached the hippocampus also took the amygdala. Without a patient whose hippocampus was damaged and amygdala spared, the series could not separate their contributions.

The postmortem study of H.M.’s brain, published in 2014 by Jacopo Annese and colleagues, measured what had actually been removed. His name, Henry Molaison, was published after his death in 2008. The resection extended 54.5 mm on the left and 44.0 mm on the right along the study’s reference axis. It took the anterior hippocampus, most of the amygdala and nearly all the entorhinal cortex, of which about 0.03 cm3 remained on the left and 0.11 cm3 on the right. About 2.02 cm3 (left) and 1.96 cm3 (right) of posterior hippocampus remained, and Nissl staining showed neuronal cell bodies in fields CA1 to CA4 and the subiculum. “Both hippocampi were removed” is therefore false.

The entorhinal cortex is the main cortical route into the hippocampus, so the remaining hippocampus may have been largely cut off. The authors proposed that this disconnection contributed more to H.M.’s amnesia than the loss of anterior hippocampus, but they offered it as a model. Living cell bodies in a section do not show the cells were connected or working. The safe statement is that H.M.’s amnesia followed a bilateral medial temporal lesion, and that his case alone cannot assign it to one structure.

Four tasks, four systems

Lesion studies in other patients and in animals linked other forms of memory to other structures. Each of these is one task in one preparation.

Striatum, probabilistic classification. Knowlton, Mangels and Squire (Science, 1996) gave a weather-prediction task in which cue cards predict one of two outcomes only probabilistically, so it is learned gradually from feedback. Amnesic patients learned it normally but had severely impaired memory for the training sessions. Nondemented patients with Parkinson’s disease, in whom dopamine input to the striatum is depleted (see the basal ganglia), failed to learn it over the first 50 trials, though they improved somewhat with further training, and remembered the training. The authors read this as the neostriatum supporting nonmotor dispositions as well as motor behaviour.

Cerebellum, eyeblink conditioning. In well-trained rabbits, a tone paired with an air puff to the eye comes to trigger eyelid closure before the puff. McCormick, Clark, Lavond and Thompson (PNAS, 1982) removed the lateral cerebellum ipsilateral to (on the same side as) the trained eye. The learned response was abolished and the reflex blink to the puff was spared. The authors suggested at least part of the plasticity might reside in the cerebellum; a lost response alone cannot separate a lost store from a cut output route. The cerebellum gives the circuit.

Amygdala, conditioned autonomic responses. Bechara and colleagues (Science, 1995) paired a stimulus with an aversive one in three patients. A patient with selective bilateral amygdala damage learned the pairing as a fact but acquired no conditioned autonomic response. A patient with bilateral hippocampal damage showed the reverse. A patient with both lesions learned neither. Emotion and stress takes up the fear side.

Why does a pair of opposite results beat one? Here is a worked problem with invented numbers in the design of the Knowlton study, not its data.

Suppose a two-choice classification task (chance 50%) and a four-choice questionnaire about the training (chance 25%). Controls score 70% and 80%. Amnesic patients score 68% and 40%. Parkinson’s patients score 52% and 78%.

Step 1: measure above chance, as a share of the controls’ gain.

relative score = (patient score − chance) / (control score − chance)

Amnesic classification: (68 − 50) / (70 − 50) = 18 / 20 = 90%. Amnesic questionnaire: (40 − 25) / (80 − 25) = 15 / 55 = 27%. Parkinson’s classification: (52 − 50) / 20 = 2 / 20 = 10%. Parkinson’s questionnaire: (78 − 25) / 55 = 53 / 55 = 96%.

Step 2: read the pattern. Each group fails the task the other passes. That is a double dissociation. A single dissociation (one group failing one task) could mean only that the failed task is harder. Here no single difficulty scale fits, because the task that is “hard” for one group is “easy” for the other.

Step 3: notice the raw-score trap. The amnesic questionnaire score of 40% against 80% looks like half of normal. Above chance it is 27%, about a quarter, because a quarter of answers are right by guessing.

Step 4: state the limits. Parkinson’s disease damages more than the striatum, and amnesic patients’ lesions differ from one another. The double dissociation shows two separable learning processes; it does not by itself pin each to one structure.

Group to edit

Amnesic patients Parkinson’s patients (nondemented)

Two-choice classification (chance 50%): 68% Four-choice questionnaire (chance 25%): 40%

View

Raw percent correct Share of the controls’ gain above chance

• Controls (fixed: 70% and 80%)
• Amnesic
• Parkinson’s
• ▲ hollow dot: above the plot, drawn at 140%; the label gives the true value

Group	Classification: raw → above-chance share	Questionnaire: raw → above-chance share
Controls
Amnesic
Parkinson’s

In this sketch: invented numbers, controls fixed at 70% and 80%, chance known exactly (50% and 25%); the picture is a sketch of the logic, not data.

Score each task above chance: relative score = (patient − chance) / (control − chance). Parkinson’s disease damages more than the striatum, and amnesic patients’ lesions differ from one another, so a double dissociation does not by itself pin each process to one structure.

Systems that interact

The four-system map is a useful specialization, not a set of sealed boxes. In male rats trained on a spatial and a cued-response maze task, Ferbinteanu (2020) found that the hippocampus supported both tasks when they were learned close in time, and the dorsolateral striatum supported both when they were learned in the same environment. Which system does the work depended on the training history.

Memories also change their dependence over time. In mice, Kitamura and colleagues (2017) found that prefrontal engram cells for a contextual fear memory (neurons active during learning whose later reactivation recalls the memory) were established during initial learning, and that by about two weeks tagged hippocampal (dentate gyrus) engram cells were no longer preferentially reactivated by natural recall cues yet could still drive recall when activated with light. Do not stretch that mouse timetable to human autobiographical memory. The cellular change underneath all of these is synaptic plasticity, and the prefrontal side of working memory is in attention and decisions. Whether new neurons in the adult human dentate gyrus contribute is debated; see adult neurogenesis.

Design problem

After a cardiac arrest a patient has memory problems, and MRI shows damage limited to field CA1 of the hippocampus on both sides. Design an assessment that decides whether the deficit is declarative only.

One way to do it

Test against matched controls, scoring above chance where tasks have chance levels.

First, separate working from long-term declarative memory: digit span (immediate) against delayed recall and recognition of a story or word list after 30 minutes. A selective declarative deficit predicts normal span and poor delayed recall.

Second, probe retrograde memory with items from recent years and from long ago, to see whether older memories are spared.

Third, test nondeclarative memory: mirror drawing across several sessions, a priming task, and probabilistic classification. A declarative-only deficit predicts normal early learning on all three (on probabilistic classification, declarative memory can matter once training runs past about 50 trials), with poor memory for the training sessions themselves.

Fourth, state the limits. A case of this kind has been reported before: patient R.B., described by Zola-Morgan, Squire and Amaral in 1986, had a marked anterograde amnesia and, at autopsy, a bilateral lesion limited to field CA1 with only minor damage elsewhere. MRI can miss damage elsewhere, so the conclusion stays provisional without postmortem histology. And one patient shows that CA1 damage can be enough for a declarative deficit; it does not show that CA1 is the declarative memory system.