Memories That Wire Themselves Together

In 1949 Donald Hebb wrote the line every neuroscience student memorises: neurons that fire together wire together. Repeated co-activation strengthens the connection. It's most of how biological memory learns what belongs with what — no supervisor, no labels, just correlation hardening into structure.

We built that as a database algorithm. Not in neural weights — in a PostgreSQL graph. When two memories are recalled together, the link between them strengthens, a little each time, easing toward a ceiling. Recall one later and its neighbours light up with it. Over time the store stops behaving like a filing cabinet and starts behaving like a mind: things that keep showing up together become genuinely, structurally associated.

Layered on top is a consolidation pass — the system's version of sleep. It distils repeated episodic memories into durable semantic ones, lets the ephemeral stuff fade, and reinforces the links that keep proving useful. What survives isn't what was stored most recently; it's what was connected most.

The result is retrieval that feels less like search and more like remembering — you reach for one thread and pull up the whole cluster it belongs to.

This one is built and running, not a proposal.

Read the full technical specification: Hebbian Co-activation in External Memory Stores →