Personal note: as @anayebi.bsky.social pointed out, this project started back at MIT during our faculty job search. This was... not an easy time. There were months when, to quote another scientist, I used to come into the office just to have the privilege of walking home with Aran 😘

P.S., the first image in this thread shows a single astrocyte in the mouse visual cortex (accessed via MICrONS: nature.com/articles/s41...), with an overlaid network of astrocyte processes and the governing energy equation for our network shown below.

Paper: pnas.org/doi/10.1073/... (6/6)

Takeaway: Memory may not live solely in synaptic weights. Our work suggests it could also reside in the molecular machinery of astrocytes. This reframes our understanding of memory—and opens new directions in neuroscience, ML, and neuromorphic computing. (5/6)

By tuning astrocyte connectivity, the model spans a spectrum: from Hopfield-like Dense Associative Memory to Transformer-like architectures. This flexibility gives rise to a continuum of highly performant systems. (4/6)

Core idea: astrocytes store information in internal calcium dynamics and inter-process communication. This astrocytic “hardware” enables many-neuron interactions and supports a superior memory scaling law—outperforming all known biologically plausible models. (3/6)

One astrocyte can interact with millions of synapses—detecting neurotransmitters, modulating synaptic strength, and signaling back to neurons. Our theory shows how these dynamics can support powerful associative memory functions. (2/6)

Ah 😁 agreed!

I agree in spirit, but is this really apples to apples? The brain was "trained" over hundreds of millions of years of biological evolution. That's a lot of lightbulbs...

Continual learning and all its attendant implications (dynamics, online control, etc)?

Da!