Glossary

Zero Data Retention (ZDR)

A model API arrangement under which your prompts and completions are not stored after the response is returned. It is a contractual promise with defined exceptions; the only technical guarantee is running the model yourself.

General definition

Zero Data Retention (ZDR) describes the strictest data-handling option that large model providers offer to API customers. Under a standard API agreement a provider typically keeps request and response content for a limited period (commonly 30 days) for abuse and misuse monitoring, and the major providers no longer train on API data by default. Under ZDR the provider commits not to persist the prompt and completion at all once the response is returned: the content is processed in memory and discarded. ZDR is granted on request, usually to enterprise customers with a reviewed use case, and is written into the contract or a data processing addendum rather than switched on by an API flag alone.

It pays to be precise about what ZDR does and does not cover:

  • Covered: prompt text, completions and usually uploaded inputs for the covered endpoints; no training on them; no retention after the request
  • Usually not covered: request metadata such as timestamps, token counts, model used and account identifiers, which are kept for billing and operations
  • Provider dependent: whether abuse monitoring still runs (in memory, without storage, or with a shorter window), which endpoints and features qualify, and whether stateful features such as stored conversations, file stores or fine-tuning are excluded because they cannot work without retention
  • Not covered at all: anything your own stack keeps: your LLM gateway logs, application logs, vector database and chat archive, which remain your responsibility

ZDR is therefore a contractual guarantee: the provider promises, and you rely on its audit reports and the DPA to hold it to that promise. It reduces the consequences of a provider-side breach and simplifies privacy assessments, but it does not change the fact that the data left your perimeter and was readable by a third party while being processed. For some data classes and some regulators that is enough; for others it is not.

The technical guarantee is self-hosting the model. An open-weight model on your own servers never sends the prompt anywhere, and retention becomes a property of your own log retention settings rather than a clause in someone else’s contract. The trade-off is operational: you run the inference hardware, choose the model and own the evaluation. Many teams combine both: a hosted model under ZDR for general tasks and a self-hosted model for the sensitive ones.

In the Ethora ecosystem

Ethora supports both routes. The AI service talks to any OpenAI-compatible endpoint, so you can point it at a cloud provider under a ZDR agreement using your own API key, or at a self-hosted model such as one served by Ollama or vLLM inside your own cloud account or data centre. In the second case the prompt, the retrieved knowledge and the reply never leave your infrastructure, which is what private AI means in practice. Switching between the two is configuration, not a rebuild.

What ZDR does not cover, Ethora puts under your control. Conversations with agents are messages in rooms, kept in the server-side archive under configurable retention with per-user erasure, so you decide how long transcripts live and can delete them on request. The compliance audit trail with immutable export records who used which agent, and on a dedicated or self-hosted deployment those records stay in the region you chose. Ethora is not itself a model provider and does not offer a ZDR contract; it gives you the architecture in which provider retention either has a contractual answer you negotiated or does not arise at all.

Get started

Chat and AI built for regulated industries

Ethora gives you audit trails, retention controls, a Trust & Safety layer and dedicated deployments in your own cloud. Talk to our team.

Start Free
Free tier available Enterprise SLA No vendor lock-in