AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems
Biology

AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems

New research shows how combining dynamic activity with temporary feature associations allows a single neural network to handle multiple cognitive demands.

By Hassan Raza
Published:
Email this Article
AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems
Photo by Sofia Lasheva on Unsplash

A new artificial intelligence model can effortlessly juggle five distinct visual memory tasks at once, shedding light on how biological brains might achieve flexible cognitive control. Researchers Mranmay Shetty and Gabriel Kreiman built a shared neural architecture that processes incoming images, categorizes distractors, and detects repetitions without requiring separate networks for each job.

The study, available on the arXiv preprint server, demonstrates that a single computational system can successfully switch between remembering a cued sample image, spotting repeating patterns, and categorizing intervening objects. Understanding how artificial systems manage these diverse cognitive demands helps scientists decode the complex mechanisms underlying human short-term memory.

Dual Memory Systems Working in Tandem

To accomplish this flexibility, the authors designed a model integrating two distinct types of temporary storage. The first is an activity state represented by a 64-dimensional vector, which captures immediate neural firing patterns that evolve as time passes. The second component consists of a dynamic associative memory matrix containing connections between image feature keys and content values.

Throughout every trial, these two memory mechanisms decay and update as new images appear. However, the underlying learned parameters of the network remain fixed. Public cues provide explicit instructions regarding the current task, allowing the same visual input to be stored and retrieved differently depending on what the task demands.

Surprising Vulnerabilities in Visual Recognition

When tested across a battery of five challenging visual tasks, the model revealed unexpected patterns in how errors occur. While blank delays of up to 256 model time units had virtually no impact on sample-matching performance, introducing distractor images and longer sequences drastically changed accuracy.

As the length of an image sequence grew from 4 items to 64 items, the false-alarm rate for detecting repetitions surged dramatically. In repetition detection tasks, the model frequently reported false familiarity when numerous novel images appeared prior to the target, even when the target itself was equally recent. Furthermore, on categorization-plus-memory tasks, the probability of complete-trial success dropped sharply as the number of distractor categories increased, falling from 76.4 percent with a single distractor down to 9.7 percent with 16 distractors.

Comparing Specialized Versus Flexible Architectures

To evaluate the unique contributions of each memory type, the researchers trained separate baseline models possessing only activity states or only associative matrices. They also tested standard sequence models, including a causal Transformer and a Gated DeltaNet architecture.

The experiments revealed that different memory types excel at different jobs. Models relying solely on activity-based memory achieved the highest performance on basic sample matching tasks, whereas models with only associative memory dominated tasks involving online repetition detection and categorization. Ultimately, the full model that combined both systems achieved robust, balanced performance across the entire suite of tasks.

The authors suggest that these computational models can generate testable hypotheses for neuroscientists recording from biological brains, particularly when comparing prefrontal cortex activity during distractor-heavy working memory tasks.

The study was posted to the arXiv preprint server on November 17, 2026.

Fact Checked

This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.

Last reviewed on .

Article history

  • Latest version

Reference(s)

  1. Shetty, Mranmay. “Déjà View: A Shared Working-Memory Model for Visual Tasks.” arXiv.org <http://arxiv.org/abs/2611.23177v1>.

Cite this page:

Raza, Hassan. “AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems.” BioScience. BioScience ISSN 2521-5760, 18 November 2026. <https://www.bioscience.com.pk/en/subject/biology/deja-view-a-shared-working-memory-model-for-visual-tasks>. Raza, H. (2026, November 18). “AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems.” BioScience. ISSN 2521-5760. Retrieved November 18, 2026 from https://www.bioscience.com.pk/en/subject/biology/deja-view-a-shared-working-memory-model-for-visual-tasks Raza, Hassan. “AI Model Juggles Five Visual Memory Tasks Using Dual Storage Systems.” BioScience. ISSN 2521-5760. https://www.bioscience.com.pk/en/subject/biology/deja-view-a-shared-working-memory-model-for-visual-tasks (accessed November 18, 2026).
End of the article