Temporal Routing

Deciding, before retrieval, which temporal question is being asked — what is true now, what was true then, or when it changed — and releasing the filters that assume the first.

Why it matters in a memory layer

A date in a question is not a matching term. "Priya is a data engineer at Northwind Labs" contains no year, so no amount of tuning the similarity function surfaces it for a question about 2025; the information the question needs lives in the validity interval, not the string. On this course's corpus the read path finds 0 of 4 such memories in its top five.

Routing alone is not enough either. A read path accumulates assumptions that the question is about the present — a belief-time filter, and a tier cap whose decay is measured from now. Releasing one gets 1 of 4; releasing both gets 4 of 4. The property that made a memory droppable is often exactly the property that makes it the answer.

Connections

Taught in: Three Temporal Questions

Used in: Retrieving Procedures

Do not confuse with: Retrieval Scoping