A vector store is not generic plumbing
Putting every embedding into an existing database looks convenient. An index, however, is defined before its model by keys, lifecycle, authority, and consumers. Conversation RAG and vault management differ on all four.
A conversation store chunks chats for answer retrieval and follows conversation lifecycle. Vestige keys by vault, path, and mirror content hash, serves curation plus authorized brain context, and reindexes from ledger commits. Shared vector shape is not a reason to share a table.
The boundary differs from Lantern too
Lantern is a corpus retrieval sibling; Vestige is a memory management surface. Turning the vault into a Lantern corpus exposes identity memory to general retrieval and creates results with no clear editing owner.
Vestige search finds source and surfaces related, duplicate, and orphan candidates for curation. Gathering context for answer generation is not its center. Similar query shape does not erase product purpose.
Mirror keys determine rebuilding
An index row needs the mirror content hash, not only a live path. A ledger commit naming a new hash triggers targeted re-embedding, while corruption permits purge and full ingest from mirror.
On rename, compare path and content identity. Distinguishing moved identical content from replacement reduces embedding spend and stale rows.
Ownership is operating responsibility
Owning an index is not preference for another database. It assigns schema migration, model version, rebuild command, health, permission filtering, and recovery.
Even separate namespaces may share credentials, retention, and destructive administration. Logical separation and ownership separation are not synonyms.