Growth Curve

Which costs move as the store grows — measured by replicating a real corpus rather than reasoned about.

Why it matters in a memory layer

Retrieval scales by sharding: more documents, more shards, a query fans out and merges. A memory layer's expensive stage compares memories with each other, so its cost is a function of how many share a comparison group, and adding shards does not shrink a group.

Measured here, eight times the store is eight times the eligible pool and 104 times the candidate pairs. The stage that breaks first is not the one users touch.

Slot blocking is doing real work and it bounds the number of groups, not the size of one: within a slot the comparison is still all-pairs, so pairs per memory climb from 0.2 to 3.2. A store with a thousand beliefs about one attribute has the problem the slot table was introduced to solve, one level down.

Replicate rather than generate, so only size varies and a change in the curve is attributable to size — and say what the harness models. Replication is many users, not one user talking for eight times as long, which matters because per-store caps multiply instead of binding.

Connections

Taught in: Scaling the Store

Do not confuse with: Partition Key