Data lifecycle
Vault storage, file versions, Trash, restore, reversible forget, audit, and retention expectations.
Stored data and index artifacts
A vault contains user-visible knowledge rows plus derived index artifacts. The stored row is the source of truth; chunks, embeddings, summaries, entities, temporal metadata, and graph links are derived so search and chat can work.
Files also keep storage metadata and version history where supported. Search artifacts can be rebuilt from the stored content by reindexing.
- Deleting a chat session does not delete source knowledge.
- Reindexing updates derived artifacts but does not change the user's original file, note, or link.
- Trash bytes still count toward storage until a row is permanently purged.
Forget and Trash
The default removal action is soft-delete. UI Move to trash and the forget_knowledge tool set forgotten metadata, remove the row from normal list/search results, and place it in vault Trash.
A single user gesture can create a forget batch. Trash groups those rows together so Restore can bring back the whole batch by using any row from that batch.
- Forgotten rows are hidden from default search, list, chat retrieval, and MCP browse results.
- The 10-second Undo toast opens Trash rather than silently restoring in the background.
- Agent and chat tools can forget; they cannot restore or hard-delete.
Restore and Delete forever
Restore is a human Trash action. It reverses forgotten rows and cascades across the forget batch so grouped actions come back together.
Delete forever is the explicit hard-purge path in the Trash UI. It is only available after a row is already forgotten, requires confirmation, and removes the row irreversibly.
- Hard purge is not exposed as an MCP or chat tool.
- Purge calls reject live rows; the lifecycle is forget first, then purge if the human confirms.
- Folder and multi-select forget operations share one batch so restore remains understandable.
Retention and audit
Workspace audit retention also controls Trash retention. When the setting is unset, the default retention window is 90 days.
A scheduled retention sweep permanently purges forgotten rows older than the workspace retention cutoff. Audit events record forget, restore, purge, and filter overrides so lifecycle changes remain traceable.
- knowledge.forgotten records via, reason, and forgotten batch.
- knowledge.restored records the batch and restored count.
- knowledge.purged records hard purge from Delete forever, retention sweep, or account-deletion flows.
Agent-safe deletion model
If an agent misreads a delete request, the worst default outcome is a recoverable move to Trash. Irreversible purge remains a deliberate human action.