Activity-Silent Working Memory via Short-Term Plasticity

28 concepts19 papers1 proposal

Traditional models of working memory rely on the concept of persistent activity: the idea that to keep a piece of information in mind, a population of neurons must fire continuously. However, empirical observations often tell a different story. In many tasks, neural firing rates return to baseline levels during the delay period, yet the information is not lost. This phenomenon has led to the theory of activity-silent working memory, where information is stored not in the electrical spikes of neurons, but in the transient physical state of the synapses connecting them.

The Synapse as a Dynamic Processor

In classical neural network theory, synapses are often treated as static "wires" with fixed weights that only change during long-term learning. Research by Si Wu 0001 et al. in Neural information processing with dynamical synapses challenges this, arguing that synapses are dynamic processors. They change their efficacy on millisecond timescales through Short-Term Plasticity (STP).

When a neuron fires, it can temporarily strengthen its connections to other neurons—a process called facilitation. As Matthijs Pals et al. explain in A functional spiking-neuron model of activity-silent working memory in humans based on calcium-mediated short-term synaptic plasticity, this facilitation is often driven by residual calcium levels in the presynaptic terminal. This "synaptic trace" allows the network to maintain a latent representation of a stimulus even after the neurons themselves have stopped firing.

Offloading Activity to Latent States

The primary advantage of activity-silent storage is metabolic efficiency. Maintaining persistent firing is energy-intensive and prone to "drift" or noise over time. By shifting the representational burden from active spiking to temporary changes in synaptic efficacy, the brain can preserve information at a much lower metabolic cost.

Jake P. Stroud et al. demonstrate in Effects of noise and metabolic cost on cortical task representations that the prefrontal cortex may use "sub-threshold dynamics" to push temporarily irrelevant information into a non-firing state. This information remains decodable from the membrane potentials or synaptic weights but does not trigger expensive action potentials. Similarly, Leo Kozachkov et al. show in Robust and brain-like working memory through short-term synaptic plasticity that STSP-enabled networks are more robust to damage and noise because the memory is embedded in the connectivity rather than the volatile hidden state.

Fast Weights and Computational Capacity

This biological insight has significant implications for artificial intelligence. Standard Recurrent Neural Networks (RNNs) and LSTMs store information in their "hidden state" (activations). However, this creates a bottleneck: the network's computational bandwidth is tied directly to its memory capacity.

Jimmy Ba proposes the use of Fast Weights to Attend to the Recent Past, introducing a third form of memory that sits between volatile activations and slow-changing weights. These fast weights act as a high-capacity associative buffer. By decoupling the network's processing state from its storage, Thomas Miconi argues in Differentiable plasticity: training plastic neural networks with backpropagation that networks can achieve much higher memorization speeds and capacities.

Maintenance vs. Manipulation

While activity-silent mechanisms are ideal for simple maintenance, they may not be sufficient for all cognitive tasks. Nicolas Y. Masse et al. (Circuit mechanisms for the maintenance and manipulation of information in working memory) found that while passive storage can be silent, the active manipulation of information—such as rotating an image in one's mind—still requires persistent neuronal firing.

This suggests a hybrid architecture where the brain (and potentially AI) switches between energy-efficient synaptic storage for "holding" data and active metabolic states for "processing" it. As Basile Confavreux et al. suggest in Memory by accident: a theory of learning as a byproduct of network stabilization, these silent traces are only revealed when the network is subsequently probed, effectively separating the latent library of the recent past from the active computation of the present.

Key Foundational Work

Neural information processing with dynamical synapses (2013) — contributed 2 concept elements to this direction.
Dynamic Modulation of Neurons and Networks (1993) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_activity_silent_working_memory_a9c5bfd9,
  title = {activity-silent working memory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/activity-silent-working-memory/rdr_a9c5bfd9b2564d8c8d1a2fcef9b980ca},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_activity_silent_working_memory_a9c5bfd9_v1,
  title = {activity-silent working memory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/activity-silent-working-memory/rdr_a9c5bfd9b2564d8c8d1a2fcef9b980ca/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake