Hard location

One entry in the codebook the memory grows for itself — an address it listens at, and the pattern it has accumulated there.

Like you're twelve

A library does not keep a shelf per book. It keeps a shelf per subject, and every book that arrives goes to the shelf it most belongs on, nudging what that shelf is about. Two hundred books might need fifty shelves. Nobody decided fifty; it is what the books turned out to need.

A hard location is a shelf. You never create one, and you never say how many.

Precisely

Each hard location holds an address — the direction it responds to — and a counter, the accumulated pattern of everything written near it. A write finds the locations nearest its input, picks a winner, adds to that winner's counter and migrates its address slightly toward the new data. The set of all locations is the codebook, and it is grown from the data rather than configured.

The count is a measurement, not a setting. Writing 1000 vectors drawn from five underlying patterns produces a few hundred locations; writing 1000 unrelated vectors produces more than a thousand, because there is more genuinely distinct material to account for. Reported as num_locations in a collection's stats.

Locations saturate. Past roughly d/32 writes a location blends what it holds, and the engine splits it rather than letting it smear — which is why capacity is counted per pool.

Related

Attractor · Associative memory · Elastic Associative Memory (EAM) · Interference