46 years ago, researchers tried treating schemas like load-bearing blueprints. It didn't work. This paper explains why and fixes it.

Mandana Samiei, Doina Precup, and Blake Richards argue in Neuron that schemas - those abstract knowledge structures that help you know what happens at a restaurant, in a classroom, or during the ritual sacrifice known as airport boarding - may not be separate mental objects at all. Instead, schemas might sit at one end of a continuum of memory abstraction, with vivid episodic memories at one end, general semantic knowledge in the middle, and highly compressed “how the world usually works” structures at the far balcony [1].

That sounds subtle. It is subtle. But subtle is where the plumbing lives.

The Old Floor Plan Had Too Many Rooms

Classic cognitive science often split memory into neat chambers. Episodic memory: the specific dinner where the waiter spilled water on your shoes. Semantic memory: knowing what a restaurant is. Schema: the general script saying you enter, sit, order, eat, pay, and quietly judge the lighting.

Architecturally, this is a tidy museum: labeled rooms, clean sight lines, no one getting lost near the gift shop. But brains are not museums. They are more like old cities: additions, alleys, patched wiring, beautiful courtyards, and one staircase that absolutely violates code.

Samiei and colleagues point out that earlier connectionist models did not require hard walls between schemas and other declarative memories. Knowledge lived in distributed patterns. The same network could store detail, gist, and abstraction depending on how experience shaped its internal structure [1]. In this view, a schema is not a special filing cabinet. It is what happens when enough related memories settle into a stable load-bearing pattern.

LLMs Walk Into the Building

The authors use modern generative AI as a useful architectural model, not because large language models have human memory, but because they show schema-like behavior without anyone installing a “schema module” next to the metaphorical break room.

LLMs learn from huge text corpora, then use context to interpret new information. When a prompt fits patterns the model already learned, it adapts quickly. When it partly fits, the model may assimilate it, sometimes squashing details into a familiar mold. When it clashes, learning requires heavier renovation, like discovering your charming bungalow is actually built on soup.

This maps neatly onto Piaget’s old terms: integration, assimilation, and accommodation. The paper’s clever move is to say: maybe these are not special schema behaviors. Maybe they are what distributed memory systems do when new information meets old structure [1].

That matters because it reframes “schema” from a thing into a level of abstraction. A detailed episode is the close-up brickwork. Semantic memory is the facade. A schema is the building code hiding underneath, quietly bossing everyone around.

Why Neuroscience Should Care

Memory research often asks where schemas live and how they differ from episodic or semantic memory. This paper suggests those questions may smuggle in the wrong blueprint. If memory forms a spectrum, then the better question becomes: how abstract is this representation, and how does that abstraction shape future learning?

Recent work fits this shift. Bein and Niv argue that schemas may emerge through reinforcement-learning principles such as prediction error, hierarchy, and dimensionality reduction in medial prefrontal cortex [2]. Reviews of memory replay show that the brain does not merely archive experiences like a dusty municipal records office. Replay can support generalization, planning, and consolidation [3]. AI researchers have also been building agents with episodic and semantic memory systems, because a model that remembers only facts but not events has the personal history of a toaster [4].

The schema-spectrum idea gives all this a cleaner load distribution. Instead of forcing memory into separate wings, it treats abstraction as a gradient across the structure.

The Real-World Payoff, If the Beams Hold

If this perspective keeps holding up, it could help AI researchers design memory systems that learn without either forgetting everything or turning every new fact into bland oatmeal. Continual learning still struggles with catastrophic forgetting, where new training overwrites old knowledge. Human memory, for all its comedy, usually avoids this. You can learn a new coffee order without forgetting your childhood address, unless the coffee is very strong.

For AI agents, the lesson is not “copy the brain.” That sentence has bankrupted enough metaphors already. The lesson is to design systems that manage abstraction: storing some details, compressing repeated patterns, and knowing when a new event should update the whole floor plan.

This is also why visual thinking tools like mapb2.io feel oddly relevant here. When you map ideas, you are doing a tiny external version of schematization: keeping some details, collapsing others, and arranging concepts so the structure becomes usable.

The Limits of the Renovation

The paper is a perspective, not a new benchmark or brain-scan result. It does not prove that schemas are “just” abstraction, and it does not claim LLMs remember like people. LLMs lack lived experience, biological consolidation, sleep replay, emotional salience, and the suspiciously specific memory of being embarrassed in seventh grade.

But the argument has elegance. It removes a fake wall. It says schemas may be less like furniture and more like architectural style: visible in many parts of the building, strongest at the abstract end, and impossible to isolate without missing the point.

For a field that loves tidy boxes, that is a useful bit of demolition.

References

[1] Samiei, M., Precup, D., & Richards, B. A. (2026). “The schema spectrum: Emergent structures and levels of abstraction in AI and the brain.” Neuron. DOI: 10.1016/j.neuron.2026.05.007. PMID: 42235490

[2] Bein, O., & Niv, Y. (2025). “Schemas, reinforcement learning and the medial prefrontal cortex.” Nature Reviews Neuroscience, 26, 141-157. DOI: 10.1038/s41583-024-00893-z

[3] Liu, Y., Dolan, R. J., Kurth-Nelson, Z., & Behrens, T. E. J. (2023). “How our understanding of memory replay evolves.” Journal of Neurophysiology, 129(3), 552-580. DOI: 10.1152/jn.00454.2022. PMCID: PMC9988534

[4] Kim, T., Cochez, M., François-Lavet, V., Neerincx, M., & Vossen, P. (2023). “A Machine with Short-Term, Episodic, and Semantic Memory Systems.” Proceedings of the AAAI Conference on Artificial Intelligence, 37(1), 48-56. DOI: 10.1609/aaai.v37i1.25075

[5] Khosla, S., Zhu, Z., & He, Y. (2023). “Survey on Memory-Augmented Neural Networks: Cognitive Insights to AI Applications.” arXiv: 2312.06141

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.