What happens to your content
API inference content is not written to a database, queue, log or trace. Saved chats are stored encrypted until you delete them.
API inference
- Content we store
- None. Prompts and outputs exist in gateway memory while the request runs.
- Protection
- TLS in transit.
- Removed
- When the request ends, apart from the idempotency cache described below.
- What stays
- Usage metadata.
Saved chat
- Content we store
- Titles, messages, reasoning, citations and attached files.
- Protection
- AES-256-GCM, with one key per conversation and one per file.
- Removed
- When you delete the conversation. Attached files are not removed with it.
- What stays
- A tombstone for the conversation, run records, tool-step labels, attached files and usage metadata.
Ephemeral chat
- Content we store
- No conversation and no messages. Encrypted run events, plus any file you upload.
- Protection
- Run events use a random key per run, held only in process memory. Uploaded files use a per-file key wrapped by the organization key.
- Removed
- Run events are deleted 1 hour after the run ends.
- What stays
- A run record, tool-step labels, any uploaded file and usage metadata.
- Idempotency cache
- If you send an Idempotency-Key on a non-streaming call, the full response stays in gateway process memory for 24 hours so a retry gets the same answer. It is never written to disk, only your organization can replay it, and a restart clears it.
- Tool-step labels
- Chat runs record each tool call's target in cleartext, with no automatic expiry: the web search query, file search query, CVE query, file name or read_url hostname. Citation URLs are cleartext too.
- Model training
- We do not use customer content to train AI models or for model analysis. Metadata is analyzed for billing, abuse monitoring and operations.
- AI data center partners
- Every partner is in the United States or China and has confirmed to NotHoly that it does not retain, store or use inference content. NotHoly verifies and checks them periodically.