Cost Profile

Where a system's model calls actually happen, counted rather than assumed.

Why it matters in a memory layer

A retrieval stack pays at read time — embed the query, search, rerank, often a call to synthesise — and indexing is a one-off amortised over every future query. That makes "index once, query forever" a sound way to think about the bill.

Measured here, the profile is inverted and starkly so: the full write path costs 48 model calls and 38 embeddings over 24 turns, and one question costs zero model calls and two embeddings. Not cheap — none.

Count calls rather than seconds. Seconds belong to the machine and the provider; calls are what a bill is denominated in and do not change when the hardware does. And count by patching the client, because a counter added at each call site misses the one somebody wrote last week.

Connections

Taught in: The Write Path Dominates

Used in: Caching, Batching, Routing · Hardening Pass · The Latency Budget · Scaling the Store

Do not confuse with: Write-Path Dominance