The Extraction Pipeline

Extraction as stages rather than one prompt: generate candidates, atomise them, gate for durability, route by type. Each stage has a single job and can be tested alone.

Why it matters in a memory layer

One prompt doing four jobs cannot be debugged — when a fact is missing you cannot tell whether it was never proposed, was merged into another, was filtered, or was mistyped. Staging also decides where cost lives: the candidate step is the only one that needs a model, so keeping the rest deterministic makes the write path cheap and auditable.

Connections

Taught in: Extraction Pipelines

Used in: The Write Path Dominates · Learning From Outcomes · Procedural Memory · Atomic Memories · Entities and Aliases · Precision and Recall on the Write Path