Minimal Data Retention

Minimal Data Retention (MDR) is available to Enterprise workspaces. It limits how long third-party model providers retain prompts and responses sent through model connections managed by Amp. It does not mean that no data is ever retained.

Under MDR, model providers retain ordinary prompts and responses only as long as needed to return a response. The following exceptions apply:

• Providers may retain safety and abuse monitoring data, including classifier results, trust and safety scores, and content flagged as potentially violating their usage policies. Each provider retains this data according to its abuse prevention and safety policies.
• Certain models or features may require longer retention. A provider may retain data for up to 30 days when specified in its current documentation.
• Providers may store computational artifacts derived from prompts, such as key/value tensors or cached tokens, for a limited period. In this case, the provider does not persist the original prompt text after generating the response.
• Providers may retain data when required by law, a court order, or a regulatory requirement.

Neither Amp nor its model providers use Enterprise customer data to train models. Model and feature eligibility for MDR may change as providers change their services and data processing terms. Some models and features are not eligible for MDR. Amp identifies these exceptions where you enable the affected model or feature.

MDR only covers data handled by third-party model providers during inference (the process of generating a model response). It does not control how Amp stores thread data. See Thread Data for Amp’s thread storage and deletion rules. If you use a customer-managed model provider connection, such as your own API key, the data retention terms of your provider account apply instead.

See Data Retention for Enterprise Subscriptions in the Terms of Service for the full legal terms.