Large-scale study of human memory for meaningful narratives

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Neuroscience

Background:

• Large-scale studies of human memory are crucial but challenging for naturalistic stimuli like narratives.
• Previous methods required extensive manual labor for stimulus design and data analysis.
• Large language models (LLMs) offer a potential solution for automating these processes.

Purpose of the Study:

• To develop and validate an LLM-powered pipeline for designing and analyzing large-scale narrative memory experiments.
• To investigate how narrative length affects recall and recognition memory.
• To explore the role of narrative comprehension and contextual reconstruction in memory.

Main Methods:

• Developed an LLM pipeline for generating narrative stimuli and analyzing memory data.
• Conducted online recall and recognition memory experiments with numerous participants.
• Utilized scrambled narratives to assess the impact of comprehension on memory.
• Employed LLM text embeddings to correlate semantic similarity with recall probability.

Main Results:

• Recall and recognition performance scaled linearly with narrative length.
• Longer narratives led to summarization rather than precise detail recall.
• Recall declined significantly for scrambled narratives, while recognition remained stable.
• Memory recall for scrambled narratives followed the original story order, suggesting contextual reconstruction.
• LLM-based semantic similarity strongly predicted recall probability.

Conclusions:

• LLMs significantly advance the scale and scope of human memory research.
• Narrative comprehension plays a critical role in recall but not recognition.
• LLMs provide new tools for memory research and psychologically informed benchmarks for AI.