Towards large language models with human-like episodic memory.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry
• Artificial Intelligence

Background:

• Episodic memory (EM) research has advanced understanding of real-world event comprehension.
• A gap exists in computational models for predicting EM use with complex, naturalistic data.
• Machine learning, particularly large language models (LLMs) with external memory, offers potential but faces alignment challenges with human memory.

Purpose of the Study:

• To review the discrepancies between current memory-augmented LLMs and human episodic memory.
• To propose criteria for benchmark tasks that align AI models with human memory.
• To outline neuroimaging methods for evaluating AI memory model predictions.

Main Methods:

• Literature review of cognitive neuroscience and AI memory research.
• Analysis of current LLM architectures and their memory augmentation techniques.
• Proposal of benchmark task criteria and neuroimaging evaluation strategies.

Main Results:

• Identified key differences between current LLM memory systems and human episodic memory.
• Established criteria for developing AI models that better reflect human memory.
• Suggested neuroimaging approaches for validating AI memory predictions.

Conclusions:

• Memory-augmented LLMs show promise but require significant alignment with human episodic memory.
• Development of standardized benchmarks and neuroimaging validation is crucial for advancing AI memory research.
• Bridging AI and cognitive neuroscience can lead to more accurate computational models of human memory.