A single inference process underlies the behavior and neural circuits of working memory

Dongyu Gong, Mario Belledonne, Ilker YildirimarXiv 2026 · arXiv

TL;DR

MemGP is a unified theory of working memory that explains behavioral limits and neural circuit dynamics through a single lens: variational inference. By treating memory as a shared, finite resource of "inducing points" in a Gaussian Process, it reconciles long-standing debates about discrete vs. continuous capacity and persistent vs. activity-silent neural states.

What Problem Does This Paper Address?

For decades, working memory (WM) research has been fragmented across different levels of analysis. At the behavioral level, researchers debate whether WM consists of a fixed number of "discrete slots" or a "continuous resource" that can be spread thin. At the neural level, there is a divide between those who believe memories are held by "persistent firing" of neurons and those who argue for "activity-silent" states (like short-term synaptic changes).

Most existing models only address one of these levels or focus on a single phenomenon, like capacity limits. The authors of this paper propose MemGP, a theory that bridges the computational, algorithmic, and circuit levels simultaneously to explain how we write, maintain, and read out memories.

Methods

The authors frame working memory as a problem of probabilistic inference. They represent memory strength as a continuous function over a feature space (e.g., location and color).

  • Computational Level: They use a Gaussian Process (GP) prior to define how memory strengths at different locations are correlated. The system performs variational inference to find a posterior distribution that balances sensory evidence against this prior.
  • Algorithmic Level: To make this computationally feasible, they use a "sparse" approximation. They represent the continuous function using a finite set of inducing points. These points are the "shared resource"—adding more items to memory doesn't add more points; instead, the existing points must be "stretched" to represent more information, naturally leading to precision loss.
  • Circuit Level: Crucially, the authors show that the math of their variational objective maps analytically onto the dynamics of a recurrent neural circuit. They don't "train" a neural network; they derive its architecture and weights directly from the inference equations.
The MemGP framework: Encoding infers a posterior, maintenance rehearses it, and retrieval samples from it.

Datasets

The study does not introduce new human data. Instead, it validates MemGP by reproducing results from several landmark working memory studies:

  • Capacity & Precision: Benchmark data from Bays et al. (2009) and Ma et al. (2014).
  • Attention/Retrocues: Data from Souza et al. (2014) and van Moorselaar et al. (2015).
  • Inter-item Biases: Data from Chunharas et al. (2022).

Key Insights and Ideas

  • Discrete Resource, Continuous Precision: MemGP reconciles the "slot vs. resource" debate. The number of inducing points is discrete and finite (like slots), but their values and positions vary continuously (like a resource). This explains why we have a capacity limit but can still represent items with varying degrees of precision.
  • Attention as Uncertainty Reduction: When a "retrocue" tells a subject to focus on one item already in memory, MemGP suggests this doesn't necessarily "strengthen" the signal. Instead, attention reweights the maintenance objective to primarily reduce the uncertainty (variance) of the cued item.
  • Emergent Connectivity: In the derived neural circuit, "center-surround" (Mexican-hat) connectivity—a staple of working memory models—emerges naturally from the GP prior's precision matrix rather than being an arbitrary design choice.
  • Hybrid Storage: The model suggests that working memory isn't just persistent activity or synaptic weights. In the derived circuit, the mean of the posterior is represented by persistent neural activity, while the covariance (uncertainty) is represented by rapidly updated "fast weights" (synaptic states).
Capacity limits emerge as items compete for a shared set of inducing points.

Key Results

  • Load-Dependent Errors: MemGP accurately reproduces the power-law decline in recall precision as the number of items (set size) increases.
  • Swap Errors: The model naturally generates "misbinding" errors (reporting the color of a non-target item) because the GP kernel correlates nearby locations.
  • Attraction and Repulsion: MemGP explains why recalled items are sometimes biased toward neighbors (attraction) and sometimes pushed away (repulsion) based on how the retrieval process normalizes the sampled posterior.
  • Neural Drift: The recurrent circuit exhibits "delay-dependent drift," where the represented feature moves away from its original value over time, matching observations in prefrontal cortex recordings.
The derived recurrent circuit reproduces the selective reduction in uncertainty following an attentional cue.

Limitations and Future Work

  • Quantitative Fitting: While the model shows excellent qualitative agreement with human behavior, the authors note that quantitative fits to individual participant data remain for future work.
  • Temporal Scaling: In the current simulations, time is measured in optimization "epochs" rather than real-world seconds. Relating these epochs to biological time scales is a necessary next step.
  • Dimensionality: The current tests are restricted to 2D spaces (location and color). Extending the model to handle multi-feature binding in higher-dimensional spaces will require more complex GP kernels.
  • Activity-Silent Recovery: While the model links uncertainty to synaptic weights, it doesn't yet demonstrate how a memory could be fully recovered if neural activity (the "mean") dropped to zero.

Summary updated

Abstract

Working memory holds past inputs online for future behavior. Despite decades of research, accounts of working memory remain divided across levels of analysis, with long-standing debates on whether its capacity reflects discrete slots or continuous resources, how attention improves recall, and whether its representations rely on persistent neural firing or activity-silent states. Here we show that MemGP, a theory that is specified at the computational, algorithmic, and circuit levels of analysis simultaneously, can explain diverse behavioral and neural phenomena in working memory. At the computational level, MemGP casts working memory as probabilistic inference over a continuous function on feature space, with a variational objective that balances memory strength against a Gaussian-process prior. At the algorithmic level, writing, storing, and reading of memories correspond to approximating this posterior with a finite set of inducing points, the resource that all items share. In simulations, MemGP reproduces load-dependent precision loss, attentional benefits during maintenance, and inter-item attraction and repulsion, in each case revealing multiple complex computational dependencies. At the circuit level, the algorithm analytically (without training) maps onto a recurrent circuit, including its architecture and dynamics. Beyond reproducing neural phenomena such as delay-dependent drift, this analytical map deduces center--surround connectivity, a property long assumed in attractor models, and a structured combination of persistent activity and synaptic weights to hold memories. Thus, a single inference objective not only jointly predicts the behavioral and neural signatures of working memory, but also addresses, across several instances and different levels of analysis, existing competing explanations.
BibTeX
@misc{gong2026single,
  author = {Dongyu Gong and Mario Belledonne and Ilker Yildirim},
  title = {{A single inference process underlies the behavior and neural circuits of working memory}},
  year = {2026},
  eprint = {2611.22054},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22054},
}
Figure