# What is Zero Data Retention?

Also called ZDR, No Retention Mode.

Zero data retention is a service configuration in which inputs and outputs are processed to fulfill a request and then not stored afterward. It is offered by some AI providers under specific plans or agreements and is frequently required in regulated procurement, though the exact scope of what is and is not retained varies considerably between vendors.

The motivation is straightforward. Prompts routinely contain customer records, source code, legal drafts and health or financial details. If a provider retains that content, the customer's risk surface now includes the provider's storage, staff access and breach history. Removing retention shortens that chain, and it removes an entire category of question from vendor review and data protection assessments.

Scope is where the detail lives. Configurations differ over whether abuse monitoring logs, safety classifier outputs, metadata such as token counts and timestamps, or error diagnostics are retained even when content is not. Some features are incompatible with zero retention by construction, including anything that must recall earlier turns, cache results, or support asynchronous retrieval of past outputs.

Zero retention on the provider side is not zero retention overall. The application still typically stores conversations to render history, indexes documents for retrieval and writes logs for debugging. Customers frequently discover that their own systems hold far more content, for far longer, than the provider they scrutinized. Retention policy therefore needs to cover the whole path, not just the model call.

Because these arrangements are contractual as well as technical, the durable version is written into the agreement rather than assumed from a settings page. Statements about retention should be verified against current provider documentation and terms, since offerings, defaults and eligibility criteria change and differ across plan tiers and regions.

## Key points

- Inputs and outputs are processed then not stored
- Abuse monitoring and metadata may still be retained
- Features needing recall or caching can be incompatible
- Application side logs and indexes often retain far more
- Confirm scope in the contract and current provider terms

## In practice

A legal team adopts an assistant for contract review and requires that clause text never persist outside their environment. They enable a no retention configuration with the model provider, then audit their own stack and find full prompts written to an application log with a one year retention. They redact those logs at write time, shorten retention to fourteen days, and document both changes in their processing record.

## Related terms

- [Training Opt Out](/en/glossary/training-opt-out)
- [Data Minimization](/en/glossary/data-minimization)
- [Data Processing Agreement](/en/glossary/data-processing-agreement)
- [Data Residency](/en/glossary/data-residency)
- [Purpose Limitation](/en/glossary/purpose-limitation)

[Back to the AI Glossary](/en/glossary)