Household memory, generation 1
The first cut at a household data layer — the platform owned the bytes behind a custom API and a forked Immich. Abandoned, and later retrodicted by the platform-formation framework: the byte layer is exactly where the abstraction inequality fails.
byte-layer abstraction · forked Immich · retrodicted by T18
Kept on the record because the abandonment is the finding.
The prototype put the platform on every app’s critical path: bytes went through a custom API, and the photo app was a fork. It worked, and it was wrong. Empirically, “the byte API was the cheap part; identity-forwarding was the cost.” In the framework’s terms, abstracting at the byte layer hides each app’s ability to use its native storage (maximal expressiveness loss, T17), the layer is expensive (a fork per app), and it destroys app self-sufficiency (T23) — so the inequality in T18 fails. The framework retrodicts the abandonment.
The stable cut is one layer up: records, not bytes. The graph of people, places and things is the abstraction; apps keep their bytes. That became Curator, the knowledge-graph service inside Nexus.