Store meaning, not tokens

A library that files by meaning, not by the words on the spine

Imagine a library where books are shelved by the words printed on the spine. Ask for “automobiles” and you’ll walk straight past a shelf of books about cars — same idea, different letters. That is how token-matching storage works: it indexes the surface text, so it only finds what spells the same, not what means the same.

Calyx’s founding premise is the opposite. Store and retrieve by meaning. An input is turned into a representation of what it is about, so “car” and “automobile” land near each other even though they share no letters.

Why a plain vector database isn’t enough

The obvious fix is a vector database: embed everything into one vector and search by similarity. That gets you meaning-based recall — but it pays for it by flattening each input into a single opaque blob. Every nuance, every distinct perspective on the input, is averaged into one point.

Calyx keeps the premise and refuses the flattening. Instead of one vector per input, it measures an input through many independent lenses and keeps each view separate. The premise — store meaning — survives; the loss of structure does not. The rest of Foundations builds up from exactly this: a lens, a slot, and then the constellation those slots form.