Why a knowledge graph beats a flat vector store

Vector stores are great for "find me text similar to this". They are terrible for "what's the current state of the answer". Drobek's knowledge graph adds three things on top of pgvector.

Temporality via supersedes edges

When you (or your agent) post a new fact that replaces an old one, you mark the old one as supersedes of the new. Search prefers the newest in a chain, but the history is still queryable for "why did we change?"

Freshness decay

Older entries get a soft penalty. If two entries are equally relevant by embedding distance, the newer one wins. The half-life is configured globally today; a per-org override is on the roadmap once anyone asks for it.

Per-org isolation enforced at the schema layer

org_id is on every row from the first migration. The withOrgScope helper enforces it in every query. There is no cross-tenant leak vector.