Zero Data Retention Policy

Last updated: November 9, 2026

This Zero Data Retention Policy (“ZDR Policy”) describes how Byteu, LLC, doing business as Fabrence (“Fabrence”, “we”, “us” or “our”), handles the content of inference requests sent to our API (the “API”). It is incorporated into our Terms of Service, and supplements our Privacy Policy.

1. Definitions

“Customer Content” means the inputs you submit to the API, such as prompts, messages, images, files and tool definitions, and the outputs the API returns to you (together, the “Inputs” and “Output” as defined in our Terms of Service).

“Zero Data Retention” or “ZDR” means that Customer Content is processed transiently, only as needed to fulfill the request, is not logged, and is not retained after processing of the request is complete, except as described in this ZDR Policy.

2. Scope

ZDR applies to every request to the API, whether you send it with a Fabrence API key or through a third-party platform we authorize (an “Authorized Channel”). It is on by default and cannot be turned off. No setting or separate agreement is required. Customers who want contractual terms may request a data processing agreement by contacting us.

3. Our commitment

For every request to the API:

Performance caches may hold intermediate model computations derived from Customer Content, as described in Section 4. We use commercially reasonable technical and organizational measures to implement ZDR. If we discover that Customer Content was retained or logged contrary to this ZDR Policy, we will delete it promptly after discovery.

4. Prompt caching

Like most inference providers, we use key-value (“KV”) caching so that requests sharing a common prefix do not have to be recomputed from scratch. A KV cache entry is an intermediate numerical representation of the model’s internal state for a portion of Customer Content. Technically, it is a compressed, model-specific encoding of that Customer Content, and we describe it here in the interest of full transparency. KV caching is standard industry practice: it is what allows work to be reused across requests and keeps inference on long contexts fast and affordable.

KV cache entries are held only in the volatile memory of the serving hardware (GPU memory and, for larger caches, host memory) and are evicted automatically after a short period of inactivity or when the memory is needed for other requests. They are never written to persistent storage, are not human-readable, are not used to reconstruct Customer Content, and are not used to train, fine-tune, evaluate or otherwise improve any model. They are used solely to serve requests to the API.

5. Data we retain

To operate, secure and bill for the API, we retain operational metadata about each request that does not include Customer Content. This includes account and API key identifiers, the model requested, timestamps, token counts, cost, latency, status and error codes, and billing records. How long we keep this metadata is described in our Privacy Policy. This ZDR Policy does not restrict our ability to retain and use such metadata, including for billing, fraud prevention, rate limiting, and service reliability purposes.

6. Exclusions

ZDR does not apply to information you send to us outside the API, such as through support requests or in account or API key names.

Fabrence is not in breach of this ZDR Policy to the extent that any logging, retention or disclosure of Customer Content results from third-party products you choose to use with the API, or from actions taken at your written direction. ZDR does not change your obligations under our Terms of Service, and you remain responsible for your use of the API and the content you submit.

7. Subprocessors

We use third-party infrastructure and model providers to operate the API (each, a “Subprocessor”). We route Customer Content only to Subprocessors that are contractually and technically configured to provide zero data retention equivalent to the commitment in Section 3. Where a Subprocessor is unable to provide such a configuration, we will not route Customer Content to that Subprocessor. Subprocessors process Customer Content only to provide services to us and are bound by written confidentiality, security and data protection obligations. A current list of Subprocessors is available to customers upon request.

Because we do not retain Customer Content, we generally cannot produce it later, including for support or legal process. If applicable law or a binding legal order requires us to preserve or disclose Customer Content, we may retain it for the affected requests to the extent required and will notify you where legally permitted.

9. Order of precedence

This ZDR Policy’s restrictions on Customer Content control over any conflicting provisions of our Terms of Service or Privacy Policy, including content licenses. A signed agreement with Fabrence may override this ZDR Policy only if it expressly identifies the provisions it overrides. The Terms of Service otherwise remain in effect, including their disclaimers of warranties and limitations of liability.

10. Changes to this ZDR Policy

We may update this ZDR Policy from time to time. When we do, we will revise the “Last updated” date above and, for material changes, notify customers through the email address on their Fabrence account or a prominent notice on our website. We will provide reasonable advance notice of any material change that narrows the commitment in Section 3. Changes apply to requests received after the updated policy takes effect.

11. Contact us

If you have questions about this ZDR Policy, please email [email protected].