Despite theoretical capacity, large language models exhibit human-like working memory limitations that worsen under cognitive load.

Pretrained LLMs struggle with working memory tasks despite transformers having full attention access to prior context, unlike simpler two-layer transformers that master these tasks perfectly

LLMs reproduce specific human working memory interference patterns: performance degrades with increased memory load and is biased by recency and stimulus statistics

Research across multiple models shows a correlation between stronger working memory capacity and broader overall competence, suggesting this is a fundamental limitation affecting AI reasoning