By default we store no prompts, no model responses, no files and no media. A request passes through the gateway in memory, goes to the provider on your own key, and streams back to you. What we keep is the record of the request: what ran, where, and what it cost.
What we keep, and for how long
- Request records — model, provider, status, timings, token counts, cost and error code, plus the
useridentifier you send, the referer and the user agent. Kept 24 months, then deleted by a monthly job. - Billing records — usage and ledger entries, kept as long as tax and accounting law requires. Once a request’s record is deleted, they no longer link to it.
- Audit log of management actions — 12 months.
- Runs — budget, counts and totals. A run’s
goaltext is cleared after 24 months. - Batches — bookkeeping only: whose batch, which provider, how many items, what it cost. Deleted once settled and the provider’s results have expired, about 30 days.
- Hosted agents — the instructions you write, encrypted, every version, and your triggers’ input templates, until you delete them. A run paused for approval holds its conversation, encrypted, until someone decides or the approval expires (24 hours by default); then it is erased.
Failed requests
A provider’s error text is returned to you in the response. We store and log only the provider, the HTTP status and the error code, because provider error text can quote your prompt.
Off unless switched on
Request logging is configured but cannot be enabled in this release (Observability export says why). The response cache stores answers only on a deployment that switches it on; GET /v1/cache says whether this one does. File uploads and video generation are off on autonomousrelay.com.
What your provider keeps
Your requests reach each provider on your own key, under your own agreement with it. Whether it keeps your prompts, and for how long, is set by that agreement and the provider’s own policy, not by us. Providers links each one’s privacy policy and terms.
Deleting what you stored
Deleting an agent deletes every version of its instructions, its triggers and the paused state of any run waiting on an approval; those runs are cancelled. Each agent’s page also has a Delete button. The full terms are in the privacy policy.