Data Retention

Everything the platform writes to its database and its buckets is either removed by something, bounded by a count, or kept on purpose. This page lists which, for the tables that grow with use. Every retention below runs on its own; a deployment sets nothing unless it wants a different age.

Removed on age

What	Kept for	Setting	Removed with it
Deleted assets, collections, feedback threads and knowledge pages	30 days after the delete	portal.deleted_retention_days	Every version object, tile and collection mosaic the item names; its shares, its places in collections and its feedback threads
Archived (deleted) memory records	90 days after the delete	memory.archived_retention_days
Records of what produced an asset or managed resource whose file is gone	90 days after the last write	portal.orphaned_producer_retention_days
Failed indexing jobs nobody resolved	90 days after they finished	apigateway.embed_jobs.failed_retention_days
Finished indexing jobs	14 days	apigateway.embed_jobs.retention_days
Records of expired webhook windows	90 days after expiry	webhooks.compactor.expired_window_retention_days
Audit log	90 days	audit.retention_days
Call catalog records nothing came of	90 days	calls.retention_days
Managed-script runs	365 days	scripts.run_retention_days

For each setting, 0 or unset takes the default shown and a negative value keeps the rows forever.

A delete in the portal hides the item at once and removes it for good once the retention has passed, so a mistaken delete can be taken up with an administrator in the meantime. A built-in knowledge page an administrator hides is never removed, because restoring the built-in pages brings it back.

The portal, memory, producer and GraphQL sweeps run once a day on every replica; each takes a PostgreSQL advisory lock of its own, so only one replica deletes at a time. The first run happens at startup. How often they run is retention.every:

retention:
  every: 24h   # default; zero or negative takes the default

Field	Type	Default	Description
retention.every	duration	24h	How often the portal, memory, producer and GraphQL sweeps run. The ages they remove at are the *_retention_days settings above
An asset whose stored object cannot be
deleted keeps its row, and the next day's run tries again, so no object is left
behind with nothing naming it.

Removed when superseded

GraphQL operation embeddings are kept for the schema a connection currently holds. The embeddings of a schema the connection has since replaced are never read again and are removed by the daily sweep. The schema itself is one row per connection.

Bounded by count

What	Bound	Setting
Asset versions	100 per asset	portal.max_versions, and per asset
Managed resource versions	10 per resource	resources.managed.max_versions
Script draft runs	20 per script and author

A pruned version's content and tiles are deleted with it.

Kept on purpose

These grow with use and have no retention, by decision:

• Script, prompt and knowledge-page version history. It is product history, the way a script's run history is: a person reading an old version or restoring one relies on it being there. Each row is the text of one version. The history goes when its script, prompt or page is removed.
• Configuration change log (config_changelog) and knowledge changesets (knowledge_changesets). They record who changed the platform's configuration and its catalog, and they grow only with administrative changes.