Analyzing Memory Effects in Large Language Models through the Lens of Cognitive Psychology

Memory, a cornerstone of human cognition, is adaptive yet fallible, as exemplified by Schacter’s seven “sins” of memory. While these phenomena are well studied in psychology and neuroscience, their presence in artificial systems, especially large language models (LLMs), is underexplored. Using paradigms from human memory research, this work systematically evaluates seven phenomena in state-of-the-art LLMs and compares them to human behavior: list length effect, list strength effect, fan effect, nonsense effect, position effect, DRM-style false memories, and cross-domain generalization. We find notable alignments: both humans and LLMs remember less under higher memory load (list length effect), benefit from repeated exposure (list strength effect), and struggle with associative interference (fan effect). LLMs also show false recognitions of semantically related but unseen items (DRM false memories) and can generalize learned associations across domains. Key divergences emerge: LLMs are less sensitive to input order (limited primacy/recency) and more robust to random or meaningless material (nonsense effect). These results clarify where LLMs echo human memory reconstruction and where architectural differences produce distinct patterns of error and resilience.

Zhaoyang Cao, Lael Schooler, Reza Zafarani