Biologically plausible implementation of key-value memory

Determine biologically plausible mechanisms and learning rules that implement key-value memory in the brain, including how keys, queries, and values are stored and associated, in a manner consistent with known neurophysiology and circuit architecture.

Background

The paper formalizes key-value memory systems where inputs are transformed into distinct key (address) and value (content) representations that are linked via an associator matrix learned with Hebbian-like updates. While association learning between keys and values has plausible biological correlates, the mappings that generate keys, queries, and values from inputs require additional biologically grounded storage and learning rules.

Developing a brain-realistic instantiation of these mappings is central to connecting machine learning architectures like transformers and fast weight programmers with neural substrates, and to explaining how hippocampus and neocortex might implement distinct roles for addressing and content storage.