Summary: New research suggests astrocytes—long considered primarily support cells—may substantially expand the brain’s memory storage capacity. Although astrocytes do not generate electrical action potentials like neurons, they influence synaptic activity through calcium signaling and release of gliotransmitters, enabling them to participate in information processing.
Using a computational framework inspired by dense associative memory, researchers propose that astrocytes can link many neurons simultaneously, dramatically increasing the number of patterns a network can reliably store and recall. The model treats individual astrocytic processes as computational subunits, yielding a memory system that is both high-capacity and energy efficient compared with neuron-only network models.
Key Points:
- Astrocytic computation: Astrocytes form tripartite synapses and use calcium-based signaling to modulate neuronal communication.
- Enhanced memory model: A neuron–astrocyte model based on dense associative memory can store far more patterns than traditional Hopfield-like networks.
- AI relevance: Insights from astrocyte-mediated computation could inspire new machine learning architectures, reconnecting modern AI with recent neuroscience findings.
Source: MIT
The human brain contains roughly 86 billion neurons. These excitable cells transmit information with electrical signals and form the backbone of conventional theories of memory storage. In addition to neurons, the brain hosts billions of astrocytes—star-shaped glial cells with many extensions that contact a vast number of synapses.
Although historically thought to perform housekeeping and metabolic support roles, a growing body of evidence indicates astrocytes participate in synaptic signaling and cognitive processes. New modeling work from MIT shows how astrocytes could play an active computational role in memory, helping to explain the brain’s enormous information capacity.
Jean-Jacques Slotine, MIT professor of mechanical engineering and brain and cognitive sciences, notes that each astrocyte can contact hundreds of thousands of synapses, making them well placed to contribute to computation rather than only support. The new open-access paper, led by Leo Kozachkov and with senior authorship by Dmitry Krotov, appears in the Proceedings of the National Academy of Sciences.
Memory capacity and astrocyte function
Astrocytes support neurons by clearing metabolic waste, supplying nutrients, and regulating blood flow. They also extend numerous fine processes that wrap around individual synapses, creating tripartite synaptic domains where presynaptic neuron, postsynaptic neuron, and astrocyte interact closely.
Recent experimental findings indicate that disrupting astrocyte–neuron connections in regions like the hippocampus impairs memory formation and retrieval. Unlike neurons, astrocytes do not fire action potentials; instead, they communicate via intracellular calcium waves and can release gliotransmitters that modulate synaptic strength and timing.
As calcium imaging techniques improved, researchers observed that astrocytic calcium dynamics correlate with nearby neuronal activity. Astrocytes detect patterns of neural firing, change their calcium levels, and respond by signaling back to synapses—forming bidirectional loops of influence between neurons and astrocytes.
What remained unclear was the computational role these calcium patterns might serve. To explore this, the MIT team built a mathematical model grounded in associative memory theory.
From Hopfield networks to dense associative memories
Hopfield networks are classical models that store and retrieve patterns through pairwise synaptic couplings. While useful conceptually, Hopfield-like networks are limited in capacity and cannot by themselves account for the massive number of memories the brain appears to hold. Dense associative memory models extend this idea by using higher-order couplings that link more than two neurons at a time, greatly increasing storage capacity.
The biological challenge is that typical synapses physically connect only two neurons. The researchers propose that astrocytes—and specifically individual astrocytic processes that contact many synapses—provide a natural biological substrate for these higher-order couplings. Because a single astrocyte can influence many synapses, a network that includes astrocytic processes can implement the multi-neuron interactions that dense associative memory requires.
Krotov explains that neuron-only networks with pairwise couplings are inherently limited in how much information they can encode. By contrast, neuron–astrocyte networks that leverage process-level astrocytic coupling can encode far richer associations, producing a substantial increase in memory capacity.
Processes as computational units
A critical innovation in the model is treating each astrocyte as a collection of independent processes rather than a single monolithic cell. Each process functions as a local computational unit: its calcium dynamics store information gradually, and gliotransmitter release communicates that information back to the specific synapses it monitors.
Because dense associative memories achieve high ratios of stored information per computational unit, the neuron–astrocyte architecture is both scalable and energy efficient. The researchers show that when astrocytic processes are included, the number of storable patterns can grow substantially with network size, leading to superior scaling compared with previously known biological implementations of dense associative memory.
Maurizio De Pitta, who was not involved in the study, highlights the striking implication: if tripartite synaptic domains serve as the brain’s fundamental computational units, then neuron–astrocyte networks could, in principle, store enormous numbers of patterns, effectively constrained by the network’s size.
Experimental tests and AI implications
To validate the model, experimentalists could develop tools to manipulate astrocytic processes and observe resulting changes in memory function. The authors hope the theoretical framework will prompt targeted experiments that probe whether astrocyte-process networks encode memory-related information.
Beyond neuroscience, the model offers conceptual guidance for artificial intelligence. By varying the connectivity among process-like units, researchers could explore a continuum of architectures ranging from dense associative memories to attention-based transformer models. Slotine argues this work helps bring modern neuroscience back into dialogue with AI, offering new architecture ideas grounded in recent biological findings.
About this memory and neuroscience research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image credit: Neuroscience News
Original research (open access): Neuron–astrocyte associative memory, Jean-Jacques Slotine et al., PNAS. DOI: 10.1073/pnas.2417788122
Abstract
Neuron–astrocyte associative memory
Astrocytes, the most abundant glial cells in the brain, play a fundamental role in memory. Despite the fact that most hippocampal synapses are contacted by astrocytes, existing theories have not yet explained how neurons, synapses, and astrocytes together support associative memory.
This work demonstrates that core aspects of astrocyte morphology and physiology naturally give rise to a dynamic, high-capacity associative memory system. The neuron–astrocyte networks described here are closely related to dense associative memory architectures used in machine learning. By adjusting connectivity patterns, the modeled family of networks spans architectures that include dense associative memories and transformers as limiting cases.
Unlike previously known biological implementations where the stored-memory-to-neuron ratio remains constant with network growth, neuron–astrocyte networks follow a superior scaling law and outperform prior biological models of dense associative memory. The model raises the novel possibility that memories might be encoded, at least in part, within networks of astrocyte processes rather than solely in synaptic weights between neurons.