Privacy and retention
What we keep, what we never see, and how to choose the shorter option.
The short version
Your prompt goes to the model. Nothing classifies it, rewrites it, or blocks it before inference, and we do not keep it as content afterwards.
What is not stored
- Prompt content. Not written to logs, not retained for review, not used to train anything.
- Response content. Same.
- Anything derived from them beyond the token counts and metadata below.
What is stored
What is needed to meter and to audit, per key:
| Item | Why |
|---|---|
| Key id, timestamp, model id | Attribution and the audit log |
| Prompt and completion token counts | Billing |
| Status code and latency | So you can see your own error rates |
| Cache hit counts | So the cached rate on your invoice is explainable |
This metadata is not content. It cannot reconstruct what you asked.
Retention
Retention applies to the metadata above, never to prompt content, because prompt content is not kept in the first place. Longer retention means a longer audit trail of your own usage, which matters if you have to show what was run and when. Where your plan lets you choose retention per key, that setting governs only that metadata; the shortest option is available on every plan.
Reserved and on-prem deployments contract zero retention of request metadata, documented in the agreement rather than in a setting.
Where processing happens
Cognitivers LLC is a United States entity. Inference runs on GPUs in the European Union, and data at rest sits in the EU. The sub-processors we rely on, and what each one sees, are listed in the data processing agreement available through contact.
Making a request about your data
For an export of the metadata held against your account, or to ask about anything on this page, use the contact address on the main site. Requests that concern a specific key should name it.
What we will not do
- We will not change the content policy of the models underneath you during a contract term without telling you, and on a reserved deployment we commit to that in writing.
- We will not read your prompts to answer a support question. If a support case needs the request, you send it to us.