For the past 50 years the representation of memory in the cerebral cortex has been the subject of continuous debate between two major theoretical positions. On one side of that debate are those who propose the subdivision of the cortex into discrete modules dedicated to special forms of memory and their specific contents. On the other side is the "holistic" position of those who propose the distribution of all memory in wide expanses of cortex. Then there are the eclectics and compromisers, who tend to take intermediate positions, such as the position that some functions or contents are localized while others are distributed. It is increasingly accepted that memory is one such function, some of its components localized in neuronal networks circumscribed to discrete domains of cortex and others widely distributed in networks extending beyond the boundaries of cortical areas defined by cellular architecture. Consequently, the aggregate of experience about oneself and the environment would be represented in cortical networks of widely ranging size and distribution. This does not exclude non-cortical structures from memory; nor does it exclude the possibility that, after acquisition in the cortex, some memories are relegated to those structures, such as the basal ganglia.
The concept of cortical network memory has two historical roots. The first is the empirical evidence, gathered in the past two centuries, that discrete lesions of the cerebral cortex rarely result in deficits of memory, while commonly affecting sensory or motor functions. Karl Lashley (1950) was the first to systematically obtain that kind of evidence by ablations of cortical areas in animals; from his observations, he inferred that memory must be widely distributed, and further, that widely dispersed neuronal assemblies represent such memories or "engrams." The second root of the concept is theoretical. Hayek (1952) was the first to formalize it in large-scale cortical networks (or "maps") that represent all experience acquired through the senses. Subsequently, artificial intelligence and connectionism further supported that concept. The essence of the concept is that mnemonic information is stored in network-like patterns of cortical connectivity established by experience. More recently, neuroscientists (Edelman and Mountcastle, 1978; Mesulam, 1998; Fuster, 2003) have come up with theoretical variants of that idea, adding to it structural and functional constraints and extending it to other cognitive functions, such as perception, which is largely a product of memory.
Here, a model of network memory in the neocortex is presented that is supported by a large body of data from neuropsychology, behavioral neurophysiology, and neuroimaging. Its essential features are:
- (a) the formation of memory by expansion of networks of cortical neurons through changes in synaptic transmission; and
- (b) the hierarchical organization of those networks, with a hierarchy of networks in parietal-temporal cortex for perceptual memory and another in frontal cortex for executive memory.
Formation of Memory
Around the previous turn of century, Cajal proposed that memory is made and stored by changes in connections between nerve cells. Subsequently, many neuroscientists expressed that notion throughout the 20th century, and for many years the notion remained accepted by many but unsubstantiated. It was theoretically formulated in considerable detail by Hebb (1949), and more recently received considerable empirical support by the acquisition of two general kinds of data (see Kandel, 1991, for review):
- the electrical generation of persistent synaptic changes in cell assemblies of the hippocampus--a phylogenetically ancient portion of cortex--and
- the induction of similar changes in the neural circuits of invertebrate animals by behavioral conditioning.
More indirect evidence indicates that synaptic plasticity or modifiability is at the basis of memory in the cerebral cortex (Fuster, 1999) and the cerebellum (Thompson, 1986). It is generally accepted that memory acquisition essentially consists in the modulation of synaptic transmission--and also, to a degree, the reduction of synapses.
Hebb (1949) postulated that two cells or systems of cells that are repeatedly activated in the cortex at the same time will become associated, so that activity in one facilitates activity in the other. Consequently, temporally coincident inputs will associate the neurons that receive them by enhancing the synapses between them. This principle, called of "synchronous convergence" (Fuster, 1999), would lead to formation of the Hebbian "cell assembly", a basic neural net of representation in the cortex. Stent (1973) made a strong theoretical argument for that principle.
Simple sensory memories can be represented in cell assemblies, nets, or modules of sensory cortex. However neuropsychologal evidence from humans and animals indicates that more complex and abstract memories and knowledge extend into areas of association cortex. Temporally coincident experiences of one or more sense modalities will modulate synapses between cells in those areas, leading to the formation of wider networks representing assorted items of individual memory, and, at higher cortical levels, of knowledge, which is the conceptual or semantic form of memory. Those larger networks extend beyond the boundaries of any given cytoarchitectonic area--numbered as per Brodmann's map in Figure 1.
The formation of the associative neuronal networks supporting -- and containing -- memory occurs along pathways of cortico-cortical connections that have been most thoroughly investigated in the nonhuman primate (Pandya and Yeterian, 1985). Those pathways depart from primary sensory and motor areas and flow into progressively higher areas of unimodal and multimodal association. The connectivity is reciprocal throughout, such that every connective step, from one area to the following, is reciprocated by fibers running in the opposite direction. At all levels and between levels, three basic features can be recognized in connective networks: convergence, divergence, and feedback or recurrence. In addition, practically all the cortical areas of one hemisphere are connected, also reciprocally, with homologous areas of the other hemisphere.
The connectivity mediating the formation of memory progresses also along maturational gradients, proceeding from area to area following the order in which the areas have myelinated in early ontogeny (Fuster, 1997). It also follows the gradients of sensory and motor processing. As memories increase in complexity, both in terms of the variety and complexity of associated experiences and the sensory inputs that convey them, the networks that sustain them become wider, and thus span progressively more associative areas of multimodal convergence.
While synaptic modulation and synchronous convergence are essential features of the process of memory formation, that process, in more general terms, is one of self-organization (Kohonen, 1984). The networks and their connectivity constitute themselves as a result of interactions of the organism with its environment. On account of these interactions, mainly by synchronous convergence, new networks are formed and old ones expanded in a dynamic process that persists through life. In the formation of memory networks, synaptic facilitation results not only from the temporal coincidence of external inputs, but also from that of these inputs with internal "inputs" resulting from the concomitant activation of preexisting parts of the network. In this manner, new memory is formed on old memory.
There has been mounting neuropsychological evidence that the hippocampus plays a critical role in the formation of neocortical memory (Squire, 1987). The hippocampus is reciprocally connected with cortical areas of association—not with primary sensory or motor areas (Amaral, 1987). This connectivity runs through and under the cortex of the peri- and ento-rhinal regions, in the medial and inferior temporal lobe. By means of that connectivity, reciprocal connections link the hippocampus with the associative cortices of the posterior (postrolandic) regions, as well as those of the frontal lobe (prerolandic), especially the prefrontal cortex. We do not know precisely the mechanisms by which the hippocampus mediates the formation of memory networks in the neocortex. Long-term potentiation (LTP) may play a role. Further, certain excitatory glutaminergic receptors, notably NMDA receptors, may also play a role. Whatever the mechanism, it results in protein changes in the membrane of cortical cells, which in turn would strengthen their synapses and thus imprint memory in cortical networks. Inputs from the amygdala, which is essential for the evaluation of the emotional and motivational significance of sensory information, may also intervene in the process.
At the interface between the environment and the cortex of association, primary sensory and motor cortices provide the inputs for the formation of memory: perceptual memory (episodic, semantic, etc.) in posterior association cortex and executive memory in frontal association—prefrontal—cortex. The primary motor cortex, in addition, provides the outputs for action from frontal association cortex (executive memory) to motor structures (pyramidal system, basal ganglia, and cerebellum). Consequently, the primary cortices—the first to myelinate in the course of development--constitute the cortical gate from sensation to memory, and from memory to action. It is appropriate, therefore, to view the structure of primary sensory and motor cortices as a form of memory at the foundation of individual memory: phyletic memory. In accord with this view, phyletic memory is constituted by the structure of primary sensory and motor cortex at birth, which is common to all organisms of the same species. Phyletic memory would be the most basic of all memories, the genetic memory that the organism has formed in the course of evolution by interactions with the surrounding world. The genetically defined structures of primary cortex dedicated to analyzing elementary sensory features and to integrating elementary primitives of movement would form the basic template on which the memory of the individual would develop. On that foundation of phyletic memory, all the memory of the individual is formed according to Hebbian principles and hierarchically organized in cortex of association. Perceptual memories organize themselves in posterior (parietal-temporal) cortex and executive or motor memories in frontal cortex.
Perceptual memory is memory acquired through the senses. This category of memory, therefore, includes a large amount of individual experience, from the simplest forms of sensory memory to the most abstract knowledge. Perceptual memories, and the networks that represent them, are hierarchically organized in the cerebral neocortex, that is, the phylogenetically “new cortex” ( Figure 1). This kind of organization develops from the bottom up, from sensory cortices to the highest association cortices. Each new perceptual memory (network) grows out of sensory cortex and, following gradients of cortical connectivity and developmental maturation, finds the level in the organization that is appropriate to its degree of concreteness or abstraction. The general schema of Figure 1 derives from a vast quantity of anatomical, neurophysiological, neuropsychological, and neuroimaging data. As the schema indicates, sensory information diverges and overlaps as it forms and is deposited at various strata of the hierarchy. Further, as the up-down bidirectional arrows indicate in the figure, the formation and retrieval of associative perceptual memory make use of connectivity between networks—i.e., memories--of different hierarchical category (heterarchical).
Phyletic memory, that is, the structure of primary sensory cortices, is at the base of the perceptual hierarchy. Above it, in cortex of sensory association, are the memories of the unimodal and polymodal sensory qualities of objects and experiences. Further up in the hierarchy, in higher associative cortex, memories are more personal and complex, with temporal and spatial associations. These memories fall under the category of what is called declarative memory, the memory of events and experiences (episodic). At the highest levels of the hierarchy, in widely distributed networks of temporal and parietal cortex called “transmodal” (Mesulam 1998), resides the knowledge of facts, concepts, and names (semantic).
At every stage of the perceptual-memory hierarchy, memory networks are formed by connections between neuronal assemblies at that stage and by convergence of inputs from below, ultimately from the senses, as well as from higher cortex. At the first stage beyond phyletic memory, networks are made of associations between sensory representations of the same modality, to form unimodal sensory memory. In polysensory association cortex, with inputs from unimodal and higher areas, more complex networks of polymodal memory are formed, associating diverse sensory inputs. Those multimodal networks constitute the basis for various kinds of episodic and semantic memory, all widely distributed and overlapping in higher areas of posterior association cortex. Networks in the cortex of areas 39 and 40 (transmodal), including the superior temporal gyrus (Wernicke's area), represent higher forms of conceptual and semantic knowledge.
Executive memory is the representation of motor acts and behaviors, widely distributed throughout the central nervous system. The spinal cord, the brain stem, and the cerebellum constitute the lowest levels of the executive hierarchy. From birth, these structures represent a large part of the motor phyletic memory of the organism. This memory is innate, stereotypical, and dedicated to the representation and fulfillment of basic drives.
The cortex of the frontal lobe represents in its networks the highest levels of executive memory, both phyletic and acquired ( Figure 1). Executive networks are formed essentially following the same principles that support the formation of perceptual networks, notably synchronous convergence. In the case of executive networks, however, the simultaneously converging inputs are of both sensory and motor origin. This includes visual and auditory stimuli that coincide with motor action and kinesthetic stimuli that accompany or result from the action; additionally, some of the inputs are made of "efferent copies" of the action itself and its components, provided by recurrent or collateral signals from the motor system (corollary discharge). Thus, to some degree executive networks in frontal cortex are the extension of perceptual networks in posterior cortex ( Figure 1, horizontal arrows).
The primary motor cortex (phyletic motor memory) is at the base of the cortical executive hierarchy. It supports the representation and execution of elementary motor acts defined by the contraction of particular muscles and muscle groups. Hierarchically above primary motor cortex is the premotor cortex (area 6), with networks representing motor acts and programs defined by goal and trajectory. These networks also participate in the structuring of language. More complex and novel programs of behavior and language are represented in the prefrontal cortex, the highest level of the executive hierarchy. The prefrontal cortex is the cortex of association of the frontal lobes. It is one of the latest neocortical regions to develop, both phylogenetically and ontogenetically (Fuster, 1997). In the human, it does not fully mature until the third decade of life. It receives profuse connections from the brain stem, limbic structures, and posterior cortex.
The executive memory networks of the prefrontal cortex represent the rules and schemas of goal-directed action. Humans with lesions of lateral prefrontal cortex have trouble remembering and making new plans. Monkeys have trouble learning and performing behavioral tasks that require the sequencing of actions, especially if the sequencing contains temporal gaps that have to be bridged by working memory (below). The so-called delay tasks (e.g., delayed response, delayed matching) exemplify this sort of behavior. Both human and nonhuman primates show disorders in it after lesions of lateral prefrontal cortex. There appears to be certain area specificity within lateral prefrontal cortex for the sensory information and motor activity that are processed in delay tasks. The importance of temporal factors, however, overrides that of regional specificity. Lesions of lateral prefrontal cortex cause deficits in the formation and execution of temporal sequences of action ("temporal gestalts"), whatever are the sensory or motor elements of those sequences.
According to imaging data, it appears that, with practice and repetition, the representation of behavioral sequences undergoes relocation from prefrontal cortex to lower levels of the motor hierarchy, especially the basal ganglia. Anterior frontal lesions in human subjects induce deficits in the performance of complex voluntary sequences without impairing automatic ones that have been relegated to hierarchically lower structures. This is the case even if these automatic movements are as complex and require as much effort to execute as when they were originally learned. Still represented in prefrontal networks after the transfer of a task to lower stages, are those aspects of the task that are subject to uncertainty or ambiguity. This is the case with delay tasks, where stimuli and responses contain both, uncertainty and ambiguity. For this reason, the correct performance of these tasks, even after thorough learning, continues to depend on the functional integrity of the lateral prefrontal cortex.
Working memory and the perception-action cycle
As noted above the formation of a memory is an associative phenomenon; so is its retrieval from permanent storage. The two phenomena are interdependent. New memory networks are formed by association between co-occurring sensory inputs, as well as between these inputs and older networks that they reactivate by association. Thus, new memory is largely the expansion of old memory. An essential point here is that every act of memory formation is accompanied by the retrieval of established memory, which is indispensable for memory formation. Probably for that reason the hippocampus plays an important role in both the acquisition and the recall of memory in the neocortex.
Neuron-activity recording in the monkey and neuroimaging in the human indicate that the retrieval of a memory consists in the activation of the cortical network that represents it (Fuster, 1999). The neuronal mechanisms of retrieval are not well understood, though they most likely involve the activation of the neuronal assemblies of a perceptual or executive network by the sensory or internal (cognitive) activation of one or several of its associated components. If the activated network has executive or motor elements, it will extend to frontal cortex, and so will its activation. If the frontal activation reaches a certain level, it will lead to action. If the action is sequential and dependent on the processing of serial perceptual inputs and motor outputs, then prefrontal cell assemblies will interact with posterior cortical assemblies in the mechanisms of the perception-action cycle and in working memory (Fuster, 1997).
The perception-action cycle is the circular flow of information that takes place between the organism and its environment in the course of a sensory-guided sequence of behavior toward a goal. Each action in the sequence causes certain changes in the environment that are analyzed bottom-up through the perceptual hierarchy and lead to the processing of further action, top-down through the executive hierarchy, toward motor effectors. These cause new changes that are analyzed and lead to new action, and so on and so forth. The cycle is thus closed at the periphery, through the environment, by sensory receptors and motor effectors. In the nervous system, the cycle is closed by connections between sensory and motor structures. Figure 2 illustrates schematically the cortical stages of the cycle. The arrows indicate connections that have been well substantiated in the primate brain; all neural connections are bidirectional. Action sequences that are automatic or well rehearsed are mediated at lower levels of the perception-action cycle. The more elaborate sequences, especially if they require the resolution of uncertainties and ambiguities in their course, are mediated at cortical levels. If a behavioral sequence contains a temporal discontinuity, as in the case of delay tasks, then the perception-action is closed at the top by working memory under control of the prefrontal cortex in co-operation with other cortical areas. That inter-area co-operation closes the temporal gap between a percept and the action that it calls for.
Working memory is the temporary active retention of information for prospective action. It is a kind of sustained attention focused on an internal representation to be used in the near term. It is essential for goal-directed behavior, speech, and reasoning. Neuropsychology and electrophysiology-- especially single-unit recording--in the primate (Fuster, 1999) have established that working memory depends on the persistent activation of a widely distributed memory network composed of associative perceptual and executive networks that have been updated and modified for an impending action. Neuroimaging in the human and nonhuman primate highlights the concomitant activation of posterior and prefrontal cortical areas during working memory. That activation presumably reflects the co-ordinated function of perceptual and executive networks in the maintenance of information at the summit of the perception-action cycle. The reciprocal anatomical connections between those networks at that highest level of the perceptual-action cycle undoubtedly play a role in that active maintenance of working memory, which may result from the reverberating reentry of impulses through those reciprocal connections.
- Amaral, D.G., 1987. Memory: Anatomical organization of candidate brain regions. In: Handbook of Physiology; Nervous System, Vol. V: Higher Functions of the Brain, Part 1, edited by Plum, F.Bethesda:Amer. Physiol. Soc., p. 211- 294.
- Edelman, G.M. and Mountcastle, V.B. The Mindful Brain. New York: Plenum
- Fuster, J.M., 1997. The Prefrontal Cortex (3rd edition), Philadelphia: Lippincott-Raven.
- Fuster, J.M., 1999. Memory in the Cerebral Cortex, Cambridge, MA: MIT Press.
- Fuster, J.M., 2003. Cortex and Mind: Unifying Cognition. New York: Oxford.
- Hayek, F.A., 1952, The Sensory Order. Chicago: University of Chicago Press.
- Hebb, D.O., 1949, The Organization of Behavior. New York: John Wiley & Sons.
- Kandel, E.R., 1991. Cellular mechanisms of learning and the biological basis of individuality In: Principles of Neural Science, edited by Kandel, E.R., Schwartz, J.H., and Jessell, T.M. Norwalk: Appleton & Lange, p. 1009- 1031.
- Kohonen, T., 1984. Self-organization and Associative Memory, Berlin:Springer.
- Lashley, K.S., 1950. In search of the engram. Symp.Soc.Exp.Biol., 4:454-482.
- Mesulam, M., 1998. From sensation to cognition. Brain, 121:1013-1052.
- Pandya, D.N. and Yeterian, E.H., 1985. Architecture and connections of cortical association areas. In: Cerebral Cortex, Vol. 4, edited by Peters, A. and Jones, E.G.New York:Plenum Press, p. 3-61.
- Squire, L.R., 1987. Memory and Brain, New York:Oxford University Press.
- Stent, G.S., 1973. A physiological mechanism for Hebb's postulate of learning. Proc.Natl.Acad.Sci.USA, 70:997-1001.
- Thompson, R.F., 1986. The neurobiology of learning and memory, Science, 233:941-947.
- Peter Redgrave (2007) Basal ganglia. Scholarpedia, 2(6):1825.
- Valentino Braitenberg (2007) Brain. Scholarpedia, 2(11):2918.
- Olaf Sporns (2007) Complexity. Scholarpedia, 2(10):1623.
- James Meiss (2007) Dynamical systems. Scholarpedia, 2(2):1629.
- William D. Penny and Karl J. Friston (2007) Functional imaging. Scholarpedia, 2(5):1478.
- Mark Aronoff (2007) Language. Scholarpedia, 2(5):3175.
- Arkady Pikovsky and Michael Rosenblum (2007) Synchronization. Scholarpedia, 2(12):1459.