Multi-dimensional, multi-scale memory during fruit fly social communication

Social communication requires continually updating one's memory of what has been said before, often over multiple timescales, yet how the brain solves this is unknown. We address this question in fly courtship, a social behavior in which the male courts the female via a long, variable song. Combining free courtship behavior and separate head-fixed neural recordings, we iteratively refined a model of female neural dynamics during courtship to resolve a novel bioplausible algorithm for the online encoding and storage of song history. Song is continually transformed by heterogeneous adaptation dynamics, increasing its dimensionality, then integrated into persistent activity, revealing how common neural processes can retain multi-dimensional, continuously unfolding information over long periods. This coding scheme produced advection-diffusion-like neural population trajectories that increased song uniqueness over multiple timescales and could be linearly transformed into flexible outputs, revealing its potential as a scalable mnemonic format poised to drive complex behavioral responses. This suggests a general mechanism for the online encoding, storage, and transformation of extended input signals, which could support a variety of behaviors.