Studying Episodic Memory in Large Language Models

‍This project investigates the capability of Large Language Models (LLMs) to mimic episodic memory, a crucial aspect of human cognition, through a series of experiments inspired by neuroscience. We designed three tasks that assess the models’ ability to recognize and utilize patterns within sequences—abilities typically reliant on episodic memory in humans. Our findings show that state-of-the-art LLMs, including LLaMA 2 70B and GPT-4, struggle with these tasks, indicating a significant limitation in their cognitive processing. We propose these tasks as a benchmark for evaluating and guiding future developments in LLMs, with the aim ofintegrating more sophisticated memory mechanisms.