arXiv study: LLMs' controllable memory makes them ideal brain research tools

In a perspective paper posted on arXiv (ID 2608.12377) and submitted to Nature Machine Intelligence, Morteza Salehjahromi, Shayan A. Zadegan, Amgad Muneer, and Jia Wu compare memory in biological brains and large language models (LLMs). Brains rely on synapses, neuronal ensembles, hippocampal-cortical interactions, and plasticity, while LLMs rely on weights, activations, context windows, retrieval systems, and external stores. Rather than mapping anatomical parts, they examine four functional questions: where memory is represented, how partial cues recover broader associations, how new information is written or updated, and how memory-related states can be perturbed. The result is a shared experimental framework for studying two very different systems.

The paper's central insight is an asymmetry in experimental access. Human studies offer observations like sparse concept responses and recall-related reactivation, but selective intervention is limited. Rodent studies provide more precise causal access, yet interventions usually affect broad circuits. LLMs, the authors argue, have no lived episodic memory but permit unusually direct, repeatable manipulation of internal states and stored information. This makes them powerful experimental instruments: researchers can precisely edit weights, rewrite activations, or stretch context windows to test theories about persistence, reversibility, and unintended side effects of memory updates. The authors stress that LLMs are not ahead in memory itself, only in the ability to perturb and inspect it. The productive path forward is transferring experimental logic—not anatomical components—from LLM research to neuroscience, turning broad questions into sharply testable biological hypotheses.