What it means
Zero data retention, or ZDR, is a setup in which an AI provider processes your request and does not keep your prompt or the model's reply afterward. Normally, providers hold some data for a while, for example in logs used to detect abuse. Under ZDR, your content is left out of those stores.
It is an agreement with the provider, not a switch every user can flip, and it is mostly a feature of business API accounts rather than consumer chat subscriptions.
Why companies ask for it
Teams handling customer records, legal files, health information or unreleased code often have rules about where data may sit and for how long. If a provider keeps nothing, there is less to leak, less to disclose in an audit and less to hand over if the provider is breached or subpoenaed. ZDR is often a requirement before a legal or security team approves an AI vendor.
What "zero" does not always cover
The phrase overstates what you get in some cases, so read the terms:
- Approval is required. Providers typically review and approve customers before enabling it.
- It can be feature by feature. Some features need to store state to work, such as saved conversations, uploaded files or stored responses, and may be excluded.
- Metadata may remain. Counts, timestamps and billing records are usually kept even when content is not.
- Law can override it. A court order can require a provider to preserve data.
- Safety systems vary. Providers differ on whether flagged requests are held for review.
ZDR is about storage after a request. It differs from training data policy, which covers whether your content is used to improve models, and from AI memory, a feature that deliberately saves details between chats. Prompt storage for speed, as in prompt caching, is a separate question worth asking your provider about, as is how an API key is protected.
What to check before you rely on it
- Get the ZDR terms in writing and read which endpoints and features are covered.
- Ask what is still logged, and for how long.
- Test your own app to confirm nothing you need depends on stored state.
- Revisit the terms when you add new features.