By 1969, we had already learned from the pioneering work of Brenda Milner that certain forms of memory were stored in the hippocampus and the medial temporal lobe. In addition, the work of Larry Squire revealed that there are two major memory systems in the brain: explicit or declarative; implicit or procedural. Explicit memory, a memory for facts and events—for people, places, and objects—requires, as Milner has pointed out, the medial temporal lobe and the hippocampus 1 2 3 . By contrast, we knew less about the localization of implicit memory, a memory for perceptual and motor skills and other forms of procedural memory which proved to involve not one but a number of different brain systems: the cerebellum, the striatum, the amygdala, and in the most elementary instances, simple reflex pathways themselves. Moreover, we knew even less about the mechanisms of any form of memory storage. Indeed, we did not even know whether the storage mechanisms were synaptic or non-synaptic.

In 1968, Alden Spencer and I were invited to write a perspective of learning for Physiological Reviews, which we entitled “Cellular Neurophysiological Approaches in the Study of Learning.” 4 In it we pointed out that there was no frame of reference for studying memory because one could not yet distinguish, experimentally, between the two conflicting approaches to the biology of memory that had been advanced up to that time: the aggregate field approach advocated by Karl Lashley in the 1950s and by Ross Adey in the 1960s, which assumed that information is stored in the bioelectric field generated by the aggregate activity of many neurons; and the cellular connectionist approach, which derived from Santiago Ramon y Cajal's idea 5 which postulated that learning results from changes in the strength of the synapse. The cellular-connection idea was later renamed synaptic plasticity by Konorski in his 1948 book: Conditioned Reflexes and Neuronal Organization 6 . Konorski’s idea was incorporated into a more specific model of certain types of learning by Hebb in 1949. Spencer and I concluded our review by emphasizing the need to develop tractable behavioral systems in which one could distinguish between these alternatives by relating, in a causal way, specific changes in the neuronal components of a behavior to modification of that behavior during learning and memory storage 4 .

The first behavioral systems to be analyzed in this manner were simple forms of learning in the context of implicit memory. From 1964 to 1979, several useful model systems emerged: the flexion reflex of cats, the eye-blink response of rabbits, and a variety of simple forms of reflex learning in invertebrates: the gill-withdrawal reflex of Aplysia, olfactory learning in the fly, the escape reflex of Tritonia, and various behavioral modifications in Hermissenda, Pleurobranchaea, and Limax, crayfish, and honeybees. The studies were aimed at defining, then pinpointing, the neural circuits that mediate these behaviors and the critical synaptic sites within these circuits that are modified by learning and memory storage, and then specifying the cellular basis for those changes 7 8 9 10 11 12 13 14 .

A number of insights rapidly emerged from this simple systems approach. The first was purely behavioral and revealed that even animals with a limited numbers of nerve cells—approximately 20,000 in the central nervous systems of Aplysia to 300,000 in Drosophila—have remarkable learning capabilities. In fact even the gill-withdrawal reflex, perhaps the simplest behavioral reflex of Aplysia, can be modified by five different forms of learning: habituation, dishabituation, sensitization, classical conditioning, and operant conditioning 15 .

The availability of these simple systems opened up the first analyses of the mechanisms of memory, which focused initially on short-term changes lasting from a few minutes to an hour. These studies found that one mechanism for learning and short-term memory evident in both the gill-withdrawal reflex of Aplysia and in the tail flick response of crayfish is a change in synaptic strength brought about by modulating the release of transmitter. A decrease in transmitter release is associated with short-term habituation whereas an increase in transmitter release occurs during short-term dishabituation and sensitization ( 16 17 18 19 20 ; for early reviews, see 21 22 ).

Studies of memory in invertebrates also delineated a family of psychological concepts paralleling those first described in vertebrates by the classical behaviorists (Pavlov and Thorndike) and their modern counterparts (Kamin, Rescorla, and Wagner). These concepts include the distinction between various forms of associative and nonassociative learning and the insight that contingency – that the conditioned stimulus, in associative learning, is predictive of the unconditional stimulus - is more critical for learning than mere contiguity: the CS preceding the US (for review see 23 ). Here, for the first time, psychological concepts, which had been inferred from purely behavioral studies, could be explained in terms of their underlying cellular and molecular mechanisms. For example, the finding that the same sensory to motor neuron synapses that mediate the gill-withdrawal reflex are the cellular substrates of learning and memory illustrates that the storage of procedural memory does not depend on specialized, superimposed memory neurons whose only function is to store rather than process information. Rather, the capability for simple procedural memory storage is built into the neural architecture of the reflex pathway.

Beginning in 1980, the insights and methods of molecular biology were being brought to bear on the nervous system, making it possible to see commonalities in the molecular mechanisms of short-term memory among different animals and to begin to explore how short-term memory is converted to long-term memory.

Even earlier, in 1974, Seymour Benzer and his students had discovered that Drosophila can learn fear and that mutations in single genes interfere with short-term memory. Moreover, flies with these mutations do not respond to either classical conditioning of fear or to sensitization, suggesting that – as in Aplysia – the two types of learning have some genes in common 8 12 . Consistent with this idea Duncan Byers, Ron Davis, and Benzer found in 1981 that in most of the mutant flies, the genes identified represented one or another component of the cAMP pathway 37 , the same pathway underlying sensitization and classical conditioning in Aplysia.

Studies of the gill-withdrawal reflex further revealed that even elementary forms of learning have distinct short- and long-term stages of memory storage. Whereas one training trial gives rise to a short-term memory lasting minutes, repeated spaced training gives rise to long-term memory lasting days to weeks 38 39 . These behavioral stages of synaptic plasticity were soon found to have parallels in the stages of the underlying synaptic plasticity – a short-term form lasting minutes to hours and a long-term form lasting days to weeks 40 41 .

The simplicity of the neuronal circuit underlying the behavioral modifications of the gill-withdrawal reflex – including direct monosynaptic connections between identified mechanoreceptor sensory neurons and their follower cells 16 19 - has allowed reduction of the analysis of the short- and long-term memory for sensitization to the cellular and molecular level. This monosynaptic sensory to motor neuron connection, which is glutamatergic 42 43 , can be reconstituted in dissociated cell culture 44 . This simplified in vitro model system reproduces what is observed during behavioral training by replacing the tail shocks with brief applications of serotonin, a modulatory transmitter normally released by sensitizing stimuli in the intact animal 26 27 28 . A single brief application of serotonin produces a short-term change in synaptic effectiveness (short-term facilitation or STF), whereas repeated and spaced applications produce changes in synaptic strength that can last for more than a week (long-term facilitation or LTF) 44 . The facilitation is also larger and longer lasting if as in classical conditioning the presynaptic sensory neuron fires action potentials just before the serotonin application, consistent with this mechanism for classical conditioning 45 46 47 .

The first clue to how short-term memory is switched to long-term memory came when Louis Flexner, followed by Bernard Agranoff and his colleagues and by Samuel Barondes and Larry Squire, observed that the formation of long-term memory requires the synthesis of new protein.

As we have seen, activation of the of Aplysia sensory neurons by serotonin or by tail stimuli leads to a local increase in cAMP and the activation of the cAMP-dependent protein kinase A (PKA) by causing the catalytic subunits of this enzyme to dissociate from the regulatory subunits. The catalytic subunits can then phosphorylate different substrates in the synaptic terminals, such as potassium channels and proteins involved in exocytosis, leading to enhanced transmitter availability and release as described above for the storage of short-term memory. When serotonin stimulation is repeated a number of times, it causes a more persistent increase in the level of cAMP that leads to longer lasting forms of synaptic plasticity. This more robust pattern of stimulation causes the catalytic subunit of PKA to recruit p42 MAPK and both then move to the nucleus where they phosphorylate transcription factors and activate gene expression required for the induction of long-term memory 48 49 .

In addition to protein kinases, synaptic protein phosphatases also play a key role in regulating the initiation of long-term synaptic changes. Various protein phosphatases, such as PP1 and calcineurin, counteract the local activity of PKA acting as inhibitory constraints of memory formation. For example, calcineurin can act as a memory suppressor for sensitization in the Aplysia 50 . Thus, an equilibrium between both kinase and phosphatase activities at a given synapse gates the synaptic signals that reach the nucleus and thus, can regulate both memory storage and retrieval 50 .

Unlike implicit memory, the conscious remembrance of things past requires a complex system involving the medial temporal lobe and the hippocampus. A new era of research was opened in 1971 when John O’Keefe made the amazing discovery that neurons in the hippocampus of the rat register information not about a single sensory modality—sight, sound, touch, or pain—but about the space surrounding the animal, a feat that depends on information from several senses (O'Keefe and Dostrovsky, 85 ). These cells which O'Keefe referred to as "place cells" fire selectively when an animal enters a particular area of the spatial environment. Based on these findings, O’Keefe and Nadel 86 suggested that the hippocampus contains a cognitive map of the external environment that the animal uses to navigate.

Independent of O’Keefe, Timothy Bliss and Terje Lømo, working in Per Andersen’s laboratory in Oslo, were also investigating the hippocampus and discovered that the synapses of the perforant pathway of the hippocampus have remarkable plastic capabilities that could serve for memory storage (Bliss and Lømo, 87 ). It soon became clear that a brief, high frequency train of action potentials in any one of the three major hippocampal pathways strengthens synaptic transmission. This long-term potentiation (LTP) has several forms. In the perforant and Schaffer collateral pathways, LTP is associative, requiring presynaptic closely followed by postsynaptic activity. In the mossy fiber pathway, LTP is nonassociative; it requires no coincident activity 88 .

A key insight into the various forms of LTP derived from Jeffrey Watkins’s discovery in the 1960s that glutamate is the major excitatory transmitter in the brain and that it acts on a number of different receptors, which he divided into two major groups: NMDA and non-NMDA (AMPA, kainate, and metabotropic) receptors. In the course of finding specific antagonists for each of these, Watkins discovered that Mg²⁺ blocks the NMDA receptor 89 . Philippe Ascher and independently Gary Westbrook next found that the Mg²⁺ blockade is voltage-dependent 90 91 . This was important because LTP in the Schaffer collateral pathway requires the NMDA receptor 89 and the receptor is unblocked when the postsynaptic cell is depolarized, which normally occurs only in response to a burst of presynaptic action potentials. Thus, the NMDA receptor has Hebbian associative properties; to release the Mg²⁺ blockade, the presynaptic neuron must be activated to provide glutamate just before the postsynaptic cell fires an action potential 89 .

Gary Lynch and Roger Nicoll next found that the induction of LTP in the Schaffer collateral pathway requires an influx of Ca²⁺ into the postsynaptic cell 92 93 . The Ca²⁺ activates directly or indirectly at least three protein kinases: (1) calcium/calmodulin protein kinase II 94 95 ; (2) protein kinase C 96 ; and (3) the tyrosine kinase fyn 97 98 .

A major question that dominated thinking in the 1980s and 1990s is whether LTP in the Schaffer collateral pathway is expressed pre- or postsynaptically. Nicoll’s finding that LTP in the Schaffer collateral pathway is associated with a selective increase in the AMPA-type receptor component of the EPSP with little change in the NMDA-type receptor component provided the first evidence that LTP at this synapse is both initiated and expressed postsynaptically 99 . Roberto Malinow and Nicoll indepdently discovered that the increase in response of the AMPA-type receptors is due to a rapid insertion of new clusters of receptors in the postsynaptic membrane from a pool of intracellular AMPA-type receptors stored in recycling endosomes 100 101 102 . Other studies, however, have implicated additional presynaptic changes that require one or more retrograde messengers from the postsynaptic cell 103 104 . These differences may depend on the frequency or pattern of stimulation used, or as suggested by Alan Fine, or on the developmental stage of the hippocampus 105 .

In 1986 Richard Morris made the first connection of LTP to spatial memory 106 by demonstrating that NMDA receptors must be activated for spatial learning in the rat. This led to a more detailed set of correlations between memory and the phases of LTP.

a) Earlier and Late LTP

Long-term potentiation in the hippocampus proved to have both early and late phases, much as long-term synaptic facilitation in Aplysia does. One train of stimuli produces the early phase (E-LTP), which lasts 1 to 3 hours and does not require protein synthesis. Four or more trains induce the late phase (L-LTP), which lasts at least 24 hours, requires protein synthesis, and is activated by PKA 107 108 .

Unlike the early phase, which can involve separate presynaptic or postsynaptic changes, the late phase depends on a coordinate structural change in both the presynaptic and postsynaptic cell through the action of one or more orthograde and retrograde messengers that assure the orderly and coordinated remodeling of both components of the synapse.

b) Molecular Similarities Between Procedural and Declarative Memory

Procedural and declarative memory differ dramatically at the behavioral and cellular level. They use a different logic (unconscious versus conscious recall) and they are stored in different areas of the brain. Nevertheless, as summarized above, the two types share in common several molecular steps and an overall molecular logic. Both are created in at least two stages: one that does not require the synthesis of new protein and one that does. In both, short-term memory involves covalent modification of preexisting proteins and changes in the strength of preexisting synaptic connections, while long-term memory requires the synthesis of new protein and the growth of new connections. Moreover, at least some examples of both forms of memory use PKA, MAP kinase, CREB-1, and CREB-2 signaling pathway for converting short-term to long-term memory. Finally, both forms appear to use morphological changes at synapses to stabilize long-term memory, and both forms require a synaptic tag 109 .

In the 1980s and 1990s, genetic analyses of behavior pioneered in Drosophila by Seymour Benzer were opened up for the mouse by Ralph Brinster, Richard Palmiter, Mario Capecci, and John Smythies (for reviews see 110 111 ). It soon became possible to selectively manipulate individual genes in an intact animal to compare the effects of such manipulations on long term memory on the one hand and on the other hand on different forms of LTP in isolated hippocampal slices. These techniques, first used to study memory by Alcino Silva in Susumu Tonegawa’s lab 112 113 and by Seth Grant in my lab 97 , revealed that interfering with LTP by knocking out specific kinases (CAMKII, fyn) also interfered with spatial memory.

Genetically modified mice were also used to determine the consequences of selective defects in the late phase of LTP. Ted Abel developed transgenic mice that expressed a mutant gene that blocks the catalytic subunit of PKA 107 . Silva and Rusiko Bourtchouladze studied mice with mutations in CREB-1 114 . Both lines of mice–those that blocked PKA with a knockout of CREB-1–had a serious defect in long-term spatial memory and both had roughly similar defects in LTP: the early phase was normal, but the late phase was blocked, providing strong evidence linking the phases of LTP to the phases of memory storage 107 114 .

In addition, a homolog of Aplysia CPEB named CPEB-3 has been found in mice raising the possibility that CPEB-3 may perform a similar function in vertebrates 115 . A parallel, self-sustaining mechanism, mediated by PKMzeta, has been discovered independently in the mammalian brain by Todd Sacktor. Blocking PKC-zeta interferes with memory even days or weeks after it is formed 116 . The finding of the roles of CPEB-3 and PKCzeta in the maintenance of memory storage suggests that in mammals as well as in Aplysia and flies memory must be actively sustained for long periods of time.

c) Consolidation and Competition in Memory

Competition between neurons is necessary for refining neural circuitry, but does it play a role in encoding memories in the adult brain? In studies of the amygdala, Sheena Josselyn and Silva found that neurons with large amounts of the CREB switch, required for long-term memory, are selectively recruited in the memory of fear. Indeed, the relative activity of CREB at the time of learning determines whether a neuron is recruited 117 . Conversely, if such neurons are deleted after learning, the memory of fear is blocked 118 .

Long-term memory requires also transcriptional activity in genes such as c-fos, zif268, that arc are rapidly and transiently induced by high frequency neural activity. Activity in these genes has therefore been used for many years to map brain activity patterns in the rodent brain. By providing a genetic readout of patterns of neural activity these genes provide the potential to obtain direct molecular control over ensembles of neurons based on their response to a given experience. In one recent study, the c-fos promoter was combined with elements of the Tet regulatory system in transgenic mice to allow the introduction of a lacZ marker into neurons activated with fear conditioning 119 . The marker provided a long-lasting record of brain activity during learning that could be compared to activity during recall. A partial reactivation of the neurons that were active during learning and the strength of the recalled memory were correlated with the degree of circuit reactivation. This important new approach provides an opportunity to introduce any genetically encoded effector molecule into neurons based on their recent activity thereby providing the potential to study circuits based on the specific memory they encode.

We now understand – in considerable molecular detail – the mechanisms underlying long-term learning-related synaptic plasticity, and the importance that such plastic changes play in memory storage, across a broad range of species and forms of memory. One surprising finding is the remarkable degree of conservation of the molecular memory mechanisms: cAMP, PKA, CRE, CREB-1 and CREB-2, and even CPEB, in different brain regions within a species and across species widely separated by evolution. In fact, one of the most striking features that has emerged through the application of molecular biology to neural science is the ability to see how unified all of the biological sciences have become.

However, although it is now clear that long-term synaptic plasticity is a key step in memory storage, it is important to note that a simple enhancement in the efficacy of a synapse is not sufficient to store a complex memory. Rather, changes in synaptic function must occur within the context of an ensemble of neurons to produce a specific alteration in information flow through a neural circuit. With the recent development of powerful genetic tools, it may soon be possible to meet the daunting challenge of visualizing and manipulating such changes in neural circuitry 120 .

It also will be interesting to see to what degree computational models will contribute to our further understanding of synaptic plasticity. The influential cascade model of synaptically stored memory by Stefano Fusi, Patrick Drew, and Larry Abbott 121 emphasizes that switch-like mechanisms are good for acquiring and storing memory but bad for retaining it. Retention, they argue, requires a cascade of states, each more stable than its precursor. As their hypothesis predicted, a progressive stabilization of changes in the synapse has been found to take place during the transition from short-term to intermediate-term to long-term memory storage (Jin et al. 122 123 ). Moreover, possible interactions between CPEB and PKC-ζ might provide additional semi-stable states within the long-term memory domain.

A major reason why computational neuroscience is rising and becoming more powerful and more interesting, as evident in the cascade model, is that these models lend themselves to experimental testing. In the future, however, computational models will need to broaden their focus to include the role of modulatory transmitters, the molecular components of synapses and their anatomical substrates.

Finally, we need to understand how memory is recalled. This is a deep problem whose analysis is just beginning. Mayford has made an important start of this problem and found that the same cells activated in the amygdala during the acquisition of learned fear are reactivated during retrieval of those memories. In fact, the number of reactivated neurons correlated positively with the behavioral expression of learned fear, indicating that associative memory has a stable neural correlate 119 . But one of the characteristics of declarative memory is the requirement for conscious attention for recall. How does this attention mechanism come into play? Do modulatory transmitters such as dopamine and acetylcholine have a role in the retrieval process?