Storage state and index resets

cacheDir contains both rebuildable indexes and durable state. The name cache.db does not mean that deleting the file is a safe index reset.

State	Ownership and recovery
Vault Markdown and attachments	Authored files; back up the vault independently.
Search chunks, embeddings and FTS/vector indexes in cache.db	Derived from readable vault content and the configured embedding representation; rebuilding may require model access and time.
Memory entities, observations and relations in cache.db	SQLite is the source of truth. Materialized Markdown is a projection, not a complete database backup.
write_provenance and write_provenance_heads in cache.db	Write-attribution history and chain anchors. Reindexing does not restore them.
Other cache.db operational state	Includes confirmations, idempotency records, captures, sessions, snapshots and event history. Preserve it unless intentionally applying its specific retention/reset policy.
experiential.db	Captured retrieval/episode history and derived signals. A fresh index does not reproduce the history. An intentional experiential reset is a separate operation.
<cacheDir>/auth.db and auth-keys/	Revocations, key retirement, signing keys and registry markers. Back them up together; registry-loss recovery is described in the auth model.
server-secrets/	Keys used for wiki seals and HITL confirmation state. Preserve them with the deployment state.

Stop all processes sharing the state directory before making a cold backup, or use a backup method that guarantees a consistent SQLite snapshot. Do not copy just a live .db file while ignoring its WAL, and do not delete database files as a shortcut for search maintenance.

Resetting the search index

Use reset_vault_cache with its confirmation flow. It is scoped to a vault and selected table families, not an entire database file. The default includes chunks, embeddings and idempotency keys; event-log and committed-capture deletion are opt-in. To preserve operational history while forcing a fresh index, explicitly set include.idempotency_keys, include.event_log and include.capture_committed to false, with include.chunks and include.embeddings set to true. Chunk deletion also clears associated chunk FTS/vector entries and may cascade to dependent derived rows. Memory entities/relations/observation history and write-provenance rows/heads are outside the tool’s reset table set.

The tool does not itself perform a reindex. Follow it with index_vault or completed boot reconciliation, and verify retrieval. For a model or dimension change, follow the coordinated embedding migration procedure, including all processes and vaults that share the database.

For cleanup of superseded embedding generations only, use include.chunks: false and include.embeddings: "inactive"; leaving chunks enabled would also delete their active embeddings. Choose the remaining flags explicitly according to the state you intend to retain.