Nearly 50 years ago, neuroscientists discovered cells in the hippocampus of the brain that store memories of specific places. These cells also play an important role in storing memories of events, known as episodic memories. Although the mechanisms by which place cells encode spatial memories have been well understood, how they encode episodic memories remains a mystery.
A new model developed by MIT researchers explains how place cells can be harnessed to form episodic memories, even without a spatial component. According to this model, place cells, along with grid cells in the olfactory cortex, act as a framework that can be used to link memories into a series of associations.
“This model is a first draft of the entorhinal cortex-hippocampus episodic memory circuitry. It provides the basis for understanding the nature of episodic memory, which is what really gets me excited,” says Ira Fiete, MIT professor of brain and cognitive sciences, a member of the MIT McGovern Institute for Brain Research, and senior author of the new study.
The model accurately simulates several features of biological memory systems, including large memory capacity, the gradual deterioration of old memories, and the ability of participants in memory competitions to store enormous amounts of information in “memory palaces.”
MIT research scientists Sarthak Chandra and Sugandha Sharma (Class of 2024) are lead authors of the study, published today in Nature . Rishdev Chaudhuri, an associate professor at the University of California, Davis, is also an author on the paper.
Index of memories
To encode spatial memories, the hippocampal place cells work in close coordination with grid cells, a specialized type of neuron that fires at different locations that are geometrically arranged in a regular pattern of repeating triangles. Collectively, a collection of grid cells forms a network of triangles that represent physical space.
These hippocampal-entorhinal circuits not only help us remember places we’ve been, but also help us navigate new places. We know from human patients that these circuits are also important for the formation of episodic memories. Episodic memories can have a spatial component, but they primarily consist of events, such as how you celebrated your birthday last month or what you had for lunch yesterday.
“The same hippocampus-entorhinal cortex circuits are used not only for spatial memory, but also for episodic memory in general,” Fiete says, “so one might ask: What is the connection that allows spatial and episodic memory to exist in the same circuits?”
Two hypotheses have been proposed to explain this functional overlap. One is that this circuit is specialized for storing spatial memories, because these types of memories, such as remembering where food is found or where predators are, are important for survival. According to this hypothesis, the circuit encodes episodic memories as a by-product of spatial memory.
Another hypothesis suggests that this circuitry is specialized for storing episodic memories, but also encodes spatial memories, since location is an aspect of many episodic memories.
In this study, Fiete and her colleagues proposed a third option, that grid cells’ distinctive tiling structure and interactions with the hippocampus are equally important for both types of memory: episodic and spatial. To develop their new model, they built on computational models her lab has developed over the past decade that simulate how grid cells encode spatial information.
“We had gotten to the point where we felt like we had some understanding of how the grid cell circuitry works, so it was time to try to understand how the grid cells interact with the larger circuitry, including the hippocampus,” Fiete said.
In the new model, the researchers hypothesize that grid cells that interact with hippocampal cells may act as a scaffold for storing spatial and episodic memories. Each activation pattern in the grid defines a “well,” and these wells are evenly distributed. Rather than the contents of a specific memory being stored in a well, each well acts as a pointer to a specific memory stored in the synapses between the hippocampus and the sensory cortex.
When a memory is later activated from the fragment, interactions between the grid and hippocampal cells bring the circuit state into the proximal well, connecting the state at the bottom of the well to the appropriate parts of the sensory cortex, which are much larger than the hippocampus and can store large amounts of memory.
“Conceptually, you can think of the hippocampus as a network of pointers. It’s like indexes that can complete patterns from partial inputs, and those indexes are sent to the sensory cortex, where those inputs are first experienced,” Fiete said. “Scaffolding doesn’t contain content, only indexes of abstract scaffolded states.”
Moreover, events occurring in succession can be linked together: each well of the hippocampal grid cell network effectively stores the information necessary to activate the next well, allowing memories to be recalled in the correct order.
Cliff and Memory Palace Model
The researchers’ new model is able to simulate several memory-related phenomena much more accurately than existing models based on Hopfield networks, a type of neural network that can store and recall patterns.
While Hopfield networks provide insight into how memories are formed by strengthening the connections between neurons, they do not fully model how biological memory works. In the Hopfield model, all memories are recalled with perfect detail until capacity is reached. At that point, no new memories can be formed, and even worse, any attempt to add more memories erases all previous memories. This “memory cliff” does not accurately mimic what happens in biological brains, where new memories are continually added while details of older memories are gradually forgotten.
The new MIT model captures insights from decades of recordings of grid cells and the hippocampus as rodents explore and forage in different environments. It also helps explain the mechanisms underlying a memory strategy called the memory palace. One of the challenges in memory competitions is to remember the order of shuffled cards in one or more decks. They often do this by assigning each card to a specific location in their memory palace – a memory of their childhood home or another familiar environment. When they need to recall cards, they walk around the house, visualizing each card in its place as they pass by. Though counterintuitive, adding the memory burden of associating cards with locations makes recall stronger and more reliable.
The MIT team’s computational model shows that the memory palace can perform such tasks very well, taking advantage of a strategy unique to memory circuits at a lower level: linking inputs to scaffolding in the hippocampus so that long-term retained memories reconstructed in the larger sensory cortex can be used as scaffolding for new memories. This allows for the sequential storage and recall of many more items than usual.
Building on this model, the researchers now plan to explore how episodic memories can be transformed into cortical “semantic” memories – memories of events that are separate from the specific context in which they were acquired (for example, Paris is the capital of France), how to define an episode, and how to integrate brain-like models of memory into modern machine learning.
This research was funded by the Office of Naval Research, the National Science Foundation’s Robust Intelligence Program, an ARO-MURI award, the Simons Foundation and the K. Lisa Yang ICoN Center.
