"Pippa should query the relevant slice of your travel life, not swallow all of it every turn."
The Dumping Trap
Waystone's whole premise is that memory compounds — so after a few years there is a lot of it. The naive way to give Pippa that memory is to stuff every past trip and every log into the prompt on every turn. That is context dumping, and it fails twice: it drowns the model in irrelevance, and it does not scale past the first handful of journeys. The right move is context retrieval — query the relevant slice, chosen by active journey, destination, season, traveler, question, linked places, and past outcomes. The goal is not to know everything at once; it is to pull exactly what this turn needs.
Two Lanes, Two Owners
Retrieval runs in two different lanes because travel memory has two different shapes:
- The structured lane. Decisive, aggregate questions — which places did Mom enjoy, which plans died to weather, fatigue against hotel moves — are answered by querying canonical Waystone state through its own tool surface. This lane is exact, and it never leaves the travel-domain store. You do not fuzzy-search a structured fact; you query it.
- The prose lane. Once assembled daily entries, decision rationales, and linked writing logs stabilize, they are captured into a corpus so consultation can cite what was actually written, with durable chunk-level provenance. This lane is for meaning and phrasing — the things that live in prose, not in columns.
Why the Split Matters
Keeping the lanes separate is what keeps each one honest. A structured question deserves an exact answer from the database, not a paraphrase that might hallucinate a date. A prose question deserves a cited passage from something Dad actually wrote, not a column value pretending to be a memory. Collapse them and you get the worst of both: fuzzy answers to precise questions and lost provenance on the prose. And there is a hard boundary the prose lane never crosses — the private memory vault is never fed into the travel corpus. Reuse means consuming a sibling's retrieval through its tool surface, not vacuuming everything into one undifferentiated index.