Models of Memory

Imagine waking up one day without any memory. You can probably picture yourself in a state of extreme confusion, not knowing who you are, where you are, or how you got there. Now, try and take things a step further. Picture waking up, looking at your bed, and not remembering what this strange object is. Everything you see would be a complete novelty to you. In this memoryless state you would be comparable to a newborn infant, guided only by basic instinct and lost without the vast repository of knowledge only gained over a lifetime. Taking in and storing information is critical to both our personal identity and our basic ability to understand the world around us. Understanding this process cues us in to the complexity of the mechanism behind memory, which is not one monolithic process but rather several intricate components. These facets come together to allow us to store a large quantity of information that we use in a wide variety of contexts.

Memory is an extremely vast topic that can be divided up in a myriad of different ways. These divisions are necessary not only to make understanding memory more manageable but also to quantify the many roles and timescales in which memory operates. The nature of memory is constantly changing; test a person’s memory five seconds, days, and years after an event and you will receive three different recollections of the same episode. In spite of its ever-shifting nature, memory can still be divided into two distinct categories: procedural and declarative. Procedural memory is memory that does not require conscious recall, common examples being the memory of how to walk or play a well-practiced instrument. Declarative memory, in contrast, presents itself in the conscious mind, allowing it to be described explicitly. The focus of this article is the latter type and the two central aspects of it, semantic and episodic memory. Semantic memory represents general facts about the world, whereas episodic memory represents specific experiences from our lives [1]. Much research has focused on creating various models explaining the mechanisms behind these two psychological phenomena as well as how they relate to each other. Beyond the drive to understand our own nature, such models are useful for understanding conditions involving memory loss and the creation of artificial systems mimicking our recall ability. Generally, the multitude of memory models can be grouped into three broad categories: psychological, physiological, and computational.

Psychological Models

The earliest models used to describe memory were psychological in nature, with the many categorizations used to differentiate memory systems tracing their origins to a psychological understanding. Within the domain of declarative memory, psychological models are chiefly concerned with describing the nature of episodic and semantic memory. The first among them to characterize these two categories of memory was described in the 1972 book, Organization of Memory, by psychologist Endel Tulving [1]. The concept of semantic memory had been in use for nearly  six years, but Tulving was the first to use the term episodic memory to describe memory that was not semantic. More important than his coinage of the new term was the clarity and specificity with which he described the two categories of memory. To this day we use his psychological model. According to his book, the fundamental difference between semantic and episodic memory is that episodic memory is tied to a specific experience whereas semantic memory has no such association. For example, everyone is aware of the fact that the sky is blue, but few people remember the specific moment in their lives when they first learned this; thus, the memory of the color of the sky is a semantic one. In contrast, consider what you ate for lunch yesterday; perhaps it was a ham sandwich. Similar to the statement that the sky is blue, saying that you ate a ham sandwich yesterday is a statement expressing a fact. However, you can most likely remember the context surrounding the latter, such as the time of day you ate the sandwich, how it tasted, and where you ate it. Thus, the memory of what you had for lunch yesterday is episodic. What sets Tulving’s model apart from contemporary understandings of memory is that psychologists at the time tended to shy away from internal phenomena that was reliant on self-reporting, opting instead to focus entirely on observable behavior. This aversion to understanding internal thoughts was to avoid speculation about phenomena that could not be directly observed [1].

Although we still have no method of directly observing someone’s memories, more advanced methods have given us insight into the physiological basis of previously abstract concepts. Tulving made no speculation about the neurobiological underpinnings of episodic and semantic memory, but the distinction between those two types of memories would later prove fundamental in understanding the neurological and computational basis of declarative memory. Tulving’s description of the intuitive distinction we can draw between the differing natures of knowledge holds a very clear physiological basis.

Physiological Models

There are two perspectives with which memory can be understood physiologically: the molecular perspective and the systems perspective. The molecular perspective mostly deals with how memory is stored [2]. Memory is not a physical object that can be moved around between different areas of the brain; instead, any change or addition to a person’s memory must occur due to a change in the properties of the neurons within the brain. How these changes come together to form the complex networks involved in semantic and episodic memory is not known, but the changes of individual neurons involved are. At the most basic level, neural networks reorganize themselves through long-term changes in synaptic connectivity strength, which quantifies how much the activation of one neuron affects the others to which it is connected. When neurotransmitters are received by a neuron, molecules classified as second messengers are released. These second messengers activate a separate group of molecules capable of altering the proteins in key areas of the neuron such as the neurotransmitter receptors or ion channels. The changes made alter the neuron’s likelihood of being activated by or activating other neurons, influencing the behavior of the network of which it is a part [2].

A systems perspective, in contrast to a molecular one, is concerned with which regions memory is stored in and how they connect. Much of what is known about the physiological basis, at a systems level, of the difference between episodic and semantic memory involves the concept of systems consolidation. This is the movement of memory between different locations in the brain. While it is not known in great detail how this transfer occurs on the small network scale, the study of retrograde amnesia has shown researchers the overarching brain areas involved [3]. Researchers had long been puzzled by why amnesia causes some memories to be lost while leaving others untouched. A potential answer to this question came from the occurrence of temporally graded retrograde amnesia, memory loss that gets worse closer to the time of the inciting incident. The observation that long-term memories were less vulnerable to disruption provided the groundwork for systems consolidation and suggested that there were two distinct ways that memories could be stored in our brains. The first model of memory involving systems consolidation described memory as being stored in two different structures of the brain, the hippocampus and the neocortex. In this model, memories were rapidly encoded into the hippocampus and other regions of the medial temporal lobe, then gradually transferred to the neocortex where they would become more stable, decreasing the chance that neurological damage would erase them. This model explains why only certain memories are lost when a person experiences amnesia. Following damage to the nervous system, the recent memories disappear, leaving the more stable consolidated memories behind [3]. This was, however, not the full picture, as the link between systems consolidation and the different types of memory was not yet identified.

The only distinction the initial understanding of systems consolidation drew between consolidated and non-consolidated memory was their levels of stability [4]. Eventually, it turned out that non-consolidated memories in the hippocampus were episodic, whereas consolidated memories in the neocortex were semantic. This finding showed that the movement of memories from the hippocampus to the neocortex not only stabilized those memories but also fundamentally altered them, “transforming” them from episodic to semantic in nature. This “transformation,” while not a direct shift of a memory from episodic to semantic, does represent a reorganization of the component sensory information from a highly contextual episodic representation to a decoupled, generalized semantic one [5]. It should also be noted that recent evidence suggests that the aforementioned component sensory information is most likely not directly stored in either the hippocampus or neocortex. Rather, the role of these two regions is more complex, dealing with the processes of encoding, consolidation, and organization of memory. For example, one study provides evidence that the lateral prefrontal cortex mediates biases affecting how stimuli trigger different compositions of working memory [6]. Nevertheless, when considering the process of systems consolidation, despite the wide distribution of memory processes, the shift that occurs can be thought of as centering on those two regions. This new model of memory “transformation” linked together the physiological and psychological understanding of memory.

The evidence for the differing physiological basis of episodic and semantic memory, like the evidence for the stability-based model of memory transformations, came from observations about memory loss. It was always known that not all amnesia was temporally graded [4]. This was a strong argument against systems consolidation until a further understanding of hippocampal damage and memory loss was reached. The critical finding was that the appearance of memory loss being temporally graded was due to only certain types of memory being lost through hippocampal damage. It was determined that semantic memories were untouched during the occurrence of temporally graded memory loss, and episodic memories were lost but not temporally graded [3].   Combined, these two facts created the perception that only older memories were lost. This showed that unconsolidated hippocampal memories were episodic and consolidated neocortical ones were semantic [4]. This new discovery opened up a wealth of questions for researchers to investigate regarding semantic memory, episodic memory, and systems consolidation. There is still much to be understood about the physiological basis of memory, but a recent understanding of the subject has allowed for another approach to answering questions about memory.

Computational Models

The most recent method of understanding memory is the development of mathematical models simulating the systems of the brain. In this perspective, human memory is viewed as a mechanism for storing and retrieving information; thus, the goal of mathematical models is to recreate this mechanism in the highly quantifiable domain of mathematics. The advantage of a mathematical understanding of neural mechanisms is twofold: Such models allow for a better understanding of the optimizing principles behind neural systems, and they provide the ability to place a computerized version of such a system in simulated environments. The optimizing principles of a system explain why it is well-suited to its current function. In the context of memory, they help explain the benefits of having multiple types of memory as opposed to a single system of information storage. For example, mathematical models can be used to demonstrate scenarios where it is useful to act based on recollection of a single recent event, without having that event drastically change the mechanism typically used for memory-based decision-making. Secondly, the ability to place a system in a simulated environment has the advantage of allowing experiments on the system to be performed without the constraints imposed by the physical world. Generally, most mathematical models of complex systems, like those found in the brain, are found in the form of computational models, mathematical models coded onto a computer.

One recent computational model, which demonstrates both of the above benefits, was used to help show that the transfer of memories from an episodic to a general context, in addition to stabilizing memories, may also improve learning in dynamic environments [7]. In the simulation, an agent, the portion of the program capable of making decisions, was tasked with finding rewards scattered throughout its environment. The agent would complete multiple foraging runs which would end every time it found a reward. In between foraging runs, when the agent was in a variable length resting period, the location of the rewards would shift. The new locations of the rewards would be related to their previously known locations in the short-term but would fall into a more general pattern unrelated to previous locations after an extended period of time. To facilitate its behavior, the agent was given both an episodic memory to recall the exact locations of previous rewards and a semantic memory to recall general patterns about reward distribution. When run through the simulation, agents with both episodic and semantic memories performed significantly better than those whose behavior was based on only a single habitual pattern [7].

Future Understandings of Memory

The understanding of semantic and episodic memory has developed considerably since their initial classification. In contrast to originally being perceived almost entirely from a psychological perspective, scientists now know much of the physiological basis of semantic and episodic memory, their relation to systems consolidation, and computational justifications for existing in the first place. With this increased knowledge has come many useful insights. For example, linking different types of memory loss to the damaged areas responsible can help us better understand and eventually treat neurological conditions, and our greater ability to computationally model memory systems can help create artificial systems capable of adapting to new information about the world [8]. Even with our comparatively advanced knowledge, there is still an enormous amount that is unknown about the way our memories function. The most glaring example of this understanding gap is that our systems-scale neurobiological models and mathematical models of memory remain relatively unconnected. We have a reasonable idea of which regions of the brain are responsible for memory, and we have several reasonable mathematical models for how a network of neurons could generate observed behavior. However, our brain imaging technology does not have the requisite combination of scale and resolution to incorporate systems-level behavior into our models of smaller neural networks. If such a model existed, it could open up an entirely new understanding of neurological conditions caused by flaws in network dynamics. Fortunately, such a reconciliation between the theorized and observed is most likely a reachable goal. If the previous pattern of discoveries is anything to go by, once the relevant data is available, it is only a matter of time before a more complete model is created.

References

1. Tulving, E. (1972) Organization of memory. New York, New York: Academic Press.
2. Kandel, E. R. (2001) The Molecular Biology of Memory Storage: A Dialogue between Genes and Synapses.    Science, 294    (5544), pp. 1030-1038.
3. Squire, L. R., & Alvarez, P. (1995). Retrograde amnesia and memory consolidation: a neurobiological perspective.    Current Opinion in Neurobiology, 5    (2), pp. 169-177.
4. Winocur, G., & Moscovitch, M. (2011). Memory Transformation and Systems Consolidation.    Journal of the International Neuropsychological society, 17    (5), pp. 766-780.
5. Battaglia, F. G., & Pennartz C. M. A. (2011). The construction of semantic memory: grammar-based representations learned from relational episodic information.    Frontiers in Computational Neuroscience, 5.  
6. Sreenivasan, K. K., Curtis, C. E., & D’Espisito, M., (2014). Revisiting the role of persistent neural activity during working memory.    Trends in Cognitive Science, 18    (2), pp. 82-89.
7. Santoro, A., Frankland, P. W., & Richards, B. A. (2016). Memory Transformation Enhances Reinforcement Learning in Dynamic Environments.    The Journal of Neuroscience, 36    (48), pp. 12228-12242.
8. Kumaran, D., Hassabis, D., & McClelland, J. L. (2016). What Learning Systems do Intelligent Agents Need? Complementary Learning Systems Theory Updated.  Trends in Cognitive Sciences 20    (7), pp. 512-534.