Standard unconstrained neural networks do not naturally separate working memory from long-term memory or impose human-like capacity limits¶

Claim · All claims
Evidence · 1 study · review r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Standard neural networks omit the working-memory/long-term-memory distinction and sharp capacity limits central to educational cognition theories, though architectural constraints can add them. → Peng Wang 2026

Evidence¶

Peng Wang 2026¶

Peng Wang, Olga Viberg, Effie Lai-Chong Law, Samuel Greiff. (2026). When Can AI Models Explain Learning? Validity Criteria for AI as Cognitive Models in Education. Educational Psychology Review. https://doi.org/10.1007/s10648-026-10202-y

q2 · i? · review · r3

Narrative review assertion about connectionist models used as virtual learners in education. The review states that without constraints, networks "can integrate and store far more information than a human learner could plausibly maintain at once", weakening their plausibility as process models, but notes this limitation is not inherent to the class.

"Most importantly, standard networks do not naturally separate short-term working memory from long-term memory, nor do they impose the sharp capacity limits that constrain human information processing"

Discussion¶

Related Claims¶

• Distinct neural substrates selectively support orthographic long-term memory and orthographic working memory in spelling — related
• Schweppe and Rummer's rejection of the multicomponent model as a basis for multimedia theories rests on a misinterpretation of the working memory–long-term memory relationship — related
• Whether working memory has a single capacity limit or multiple capacity limits from functional specialization is highly relevant for the design of multimedia materials — related
• Chunking reduces working memory load by grouping information into fewer, more meaningful units. — related