Agent records and self-scheduling
Use persistent records and scheduled follow-ups within an agent’s allowed capabilities.
In this topic
An agent database stores deliberate records: a reading queue, review findings, an inventory, or open decisions. Self-scheduling lets that agent request a future run within configured limits. You can enable either feature independently.
Memory is different. It extracts conversational context; a database gives the agent tables and explicit operations. Choose the database when you need fields, filtering, exports, and an auditable change history.
Enable the database
Open the agent's abilities and enable Database. Give it a small schema-oriented task:
Keep a reading queue with title, source URL, status, and date added.
Add the two articles I provide. Do not infer that an article is read
unless I say so. Show the current queue after each update.
Inspect the created table before relying on it. A generated statement that a record was saved should be backed by a query or the database view.
Review records
| View | Use |
|---|---|
| Overview | Pinned saved views for quick inspection |
| Tables | Rows, columns, filtering, edits, and exports |
| Saved Views | Reusable queries with meaningful names |
| History | Changes associated with recorded runs |
Rows deleted through the agent tools are soft-deleted. Active, Deleted, and All filters help distinguish a hidden row from a permanently removed record. Restoration is available through the corresponding database operation.
Database operations
The tool surface includes schema inspection and changes, inserts, updates, upserts, deletes, restores, queries, saved views, and import/export. db_query is a bounded read path; inspect truncation when querying a large table. db_execute is subject to host policy and should not be treated as unrestricted SQL access.
Created records carry reserved timestamps such as _created_at, _updated_at, and _deleted_at. The change log connects mutations to runs. Do not use those internal columns as ordinary application fields.
The default storage limit is 100 MiB per agent, with a warning threshold of 80%. Soft deletion does not necessarily shrink the physical database file immediately. If quota checks reject a write, review storage and maintenance options rather than repeatedly deleting rows and assuming the file has become smaller.
Privacy and persistence
The database is stored with Mellow's local data. A remote model can receive schema information and any rows read into its request context. Keeping the file locally does not keep all queried content out of provider requests.
Disabling the ability prevents normal agent use but does not mean the stored data has been erased. Use the appropriate data-management action when removal is your intent. Encryption follows the configured storage stack; see Storage.
Self-scheduling
Enable Self-scheduling separately. The agent can set one pending wake-up with a brief describing the next task. A run clears that slot before dispatch; continued recurrence requires another scheduling request.
A scheduled responsibility you create in the Schedules screen and an agent's next-run slot are different mechanisms. Use the former for a fixed recurring brief, and the latter for a bounded follow-up that depends on the current work.
Schedule modes
| Mode | Horizon | Minimum interval | Daily cap | Quiet hours |
|---|---|---|---|---|
| Ambient | 7 days | 1 hour | 6 | 22:00–07:00 |
| Reactive | 24 hours | 5 minutes | 48 | None |
| Project | 30 days | 1 hour | 4 | 22:00–07:00 |
Manual is the off state. Mode selection applies a preset, not just a display label. Requests outside a mode's bounds can be adjusted to an allowed time, with an explanation returned to the agent.
Next-run request
schedule_next_run accepts either an absolute scheduled_at time or relative in_seconds, plus required instructions. Optional context_views identify saved views to prefetch. on_miss selects skip, run_once, or run_catchup behavior for an overdue time. priority records normal or low intent; do not assume it guarantees interruption behavior.
cancel_next_run clears the pending slot. notify produces a user-facing notification through its supported path. These operations are unavailable when the scheduling ability is disabled.
Pause and diagnose
Use the Next Run controls to pause for a duration or indefinitely. Check the saved next-run value after changing mode or resuming. If no follow-up arrives, inspect the ability toggle, pause state, miss policy, app availability, and the run's model/tool errors.
For records that must be durable outside the application, periodically export and inspect them. A scheduled notification should refer to the saved record rather than serve as its only copy.
Continue exploring · Connect your workspaceDevices and Mobile →Pair a client with a host Mac and understand where its conversations and tools run.