[AIP Distinguished Lecture] Talk by Prof. Haim Sompolinsky (Harvard University and Hebrew University) on "Lifelong Memory in Large Language Models"

A RIKEN AIP Distinguished Lecture by Prof. Haim Sompolinsky (Harvard University and Hebrew University) on lifelong memory in large language models. LLMs mostly take in new information through context windows and retrieval-augmented generation, while encoding experiences directly into the weights through sequential fine-tuning tends to suffer from catastrophic forgetting. The talk presents an approach based on the complementary learning systems framework from cognitive neuroscience: a fast, hippocampus-like system in which Gated LoRA stores each experience in its own sparse low-rank adapter, and a slow, cortex-like system in which replay consolidates memories into the base weights. It also introduces a statistical-mechanical theory of memorization through low-rank fine-tuning, covering how memory capacity scales with trainable parameters and how pretrained knowledge improves recall. The lecture is hybrid: it is streamed online, and in-person attendance at the RIKEN Nihonbashi Office is limited to AIP researchers.