# Zero data retention

> What Kurrens stores about your requests — and what it never stores.

Kurrens runs open-weight models on infrastructure we operate in Johor, Jakarta, and Dublin (see
[Regions](/docs/getting-started/regions)). In every region, your prompts and completions are processed in memory for the
length of the request and then discarded.

| | Prompts & inputs | Completions & outputs | Request metadata |
|---|---|---|---|
| Stored on disk | No | No | Yes |
| Written to logs | No | No | Yes |
| Used to train or improve models | No | No | No |
| Shared with model developers | No | No | No |
| Retention | None | None | 13 months |

Zero data retention applies to **every account by default**. It can't be turned off, and we don't offer discounts in
exchange for your data.

## Request metadata

To bill and secure the service we record, per request: request id, timestamp, API key id, model id, service tier, token
counts, latency, status code, client IP address, and cost. Metadata never contains the content of prompts or completions.

## The only exceptions

1. **Support you ask for.** If you ask us to debug a specific request and give written permission, we capture those
   requests only, and delete them within 30 days after the issue is resolved.
2. **Legal requirements.** If the law requires us to preserve specific data, we do only what's required and, where
   allowed, tell you.
3. **Batch jobs** (where offered) are stored encrypted until results are delivered, then deleted.

## Prompt caching

Cached model state stays in memory, is scoped to your account, is never written to persistent storage, and expires
automatically. See [Prompt caching](/docs/inference/prompt-caching).

The binding text is our [Data Policy](/legal/data-policy), which takes precedence over any other policy
we publish.