Towards large language models with human-like episodic memory (2025)
Cognitive neuroscience has made significant strides in understanding how episodic memory (EM), our memory for unique past experiences, helps us comprehend real-world events. However, a major challenge is the lack of computational models that can predict how EM is used with high-dimensional, naturalistic stimuli. Memory-augmented large language models (MA-LLMs) are presented as a promising framework because they combine rich semantic knowledge, similar to human semantic memory, with an external memory store analogous to EM. Despite this potential, the paper argues that current MA-LLMs are misaligned with human EM in several fundamental ways. Key differences include that human memory is dynamically updated and even altered after initial storage, whereas MA-LLMs typically store memories in a fixed, unchangeable format. Furthermore, humans automatically segment continuous experiences into discrete events, which influences memory encoding and retrieval, a feature largely absent in the fixed-size chunking used by most models. Humans also exhibit selective memory processes—encoding and retrieving information more strongly at specific moments, like at event boundaries or when understanding fails—and demonstrate temporal contiguity, where recalling one memory primes the recall of temporally adjacent ones, properties that are not standard in MA-LLMs.
To bridge the gap between machine and human memory, the authors propose developing new benchmark tasks that better reflect the complexities of real-world memory use. Unlike current question-answering tasks with clear prompts, these proposed benchmarks would involve continuous information streams, ambiguity about when to store or retrieve memories, and the risk of retrieving incorrect information from a cluttered memory store. Such challenges, the paper suggests, would encourage the development of MA-LLMs with more human-like features, such as selective retrieval policies and competitive memory access. The authors also advocate for rigorous evaluation of these models against human data, using neuroimaging techniques to compare the models' internal representations with brain activity during naturalistic tasks like listening to stories or watching movies. By aligning MA-LLMs more closely with the architectural properties of human EM, the authors believe it will not only advance cognitive modeling, making it possible to predict memory use for novel stimuli, but also enhance the capabilities of AI systems by incorporating the efficient and adaptive strategies of human memory.
