---
title: Which LLM API features are excluded from zero data retention?
url: https://opencontroller.lyzr.com/blog/llm-api-features-excluded-from-zero-data-retention
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# Which LLM API features are excluded from zero data retention?

**Zero data retention (ZDR) removes your prompts and outputs from the provider's abuse-monitoring logs for stateless calls. It does not cover features whose job is to store state: files, vector stores, threads, conversations, batches, fine-tuning and evals. If an agent in a ZDR project calls those endpoints, its data is stored until deleted. Block those endpoints for ZDR agents at the gateway.**

Security teams often sign a ZDR amendment and assume every agent built on that provider is covered. Agents are exactly where the assumption breaks, because agent frameworks lean on stateful features: hosted conversation history, file search, code execution containers and background jobs.

This guide uses OpenAI's published data-controls table as the worked example, because it lists retention per endpoint. Other providers publish similar exclusions; check yours.

## What does zero data retention actually remove?

OpenAI keeps API abuse-monitoring logs for up to 30 days by default. Two controls change that, and both need prior approval from OpenAI:

- **Modified Abuse Monitoring (MAM)** excludes customer content from abuse-monitoring logs on all endpoints (image and file inputs can still be retained in rare cases).
- **Zero Data Retention (ZDR)** does the same and also changes endpoint behaviour. For example, the `store` parameter on `/v1/responses` is always treated as `false`, even if the request sets it to `true`.

Both can be applied to a whole organisation or to individual projects. That matters: a ZDR organisation can still contain a project that does not inherit the setting.

What ZDR does not do is delete **application state** — data you asked the service to keep so a feature can work.

## Which OpenAI endpoints are not ZDR-eligible?

From OpenAI's data-controls documentation:

| Endpoint | ZDR eligible | Application state kept |
|---|---|---|
| `/v1/chat/completions`, `/v1/responses` | Yes, with limits | None, with exceptions |
| `/v1/embeddings`, `/v1/completions`, `/v1/realtime` | Yes | None |
| `/v1/conversations`, `/v1/chatkit/threads` | No | Until deleted |
| `/v1/assistants`, `/v1/threads`, `/v1/vector_stores`, `/v1/agents` | No | Until deleted |
| `/v1/files` | No | Until deleted or expired |
| `/v1/batches`, `/v1/fine_tuning/jobs`, `/v1/evals` | No | Until deleted |
| `/v1/images/generations` | Yes, with limits (specific models only) | None |

Three agent-specific details are easy to miss:

1. **Assistants objects persist.** Objects related to the Assistants API are deleted 30 days after you delete them. If you never delete them, they stay.
2. **Background mode** on the Responses API keeps data on disk for about 10 minutes so the client can poll for the result.
3. **Code interpreter containers** hold state while active and are deleted when they expire or you delete them.

## Why do agents hit these exclusions more than chat apps?

A chat app usually sends a stateless completion and stores history in your own database. Agent frameworks often outsource that state to the provider:

- **File search** needs an uploaded file and a vector store — both not ZDR-eligible.
- **Hosted conversation state** (`/v1/conversations`) keeps every turn until deleted.
- **Evaluation pipelines** upload datasets to `/v1/evals` or `/v1/files`.
- **Batch jobs** for overnight document processing use `/v1/batches`.

A developer can enable any of these with one line of SDK code. Nobody in security sees the change unless something is watching the traffic.

## How do you keep ZDR agents away from stateful features?

Use this checklist for every agent that handles data covered by your ZDR agreement:

1. **Confirm the project setting.** Check that the agent's API project has ZDR applied, not just the organisation. Record the project ID in the agent's registry entry.
2. **Tag the agent as ZDR-required** in your inventory, with the data class that drives the requirement.
3. **Allowlist endpoints at the gateway.** For ZDR agents, allow only eligible paths such as `/v1/responses` and `/v1/embeddings`. Block `/v1/files`, `/v1/vector_stores`, `/v1/assistants`, `/v1/threads`, `/v1/conversations` and `/v1/batches`.
4. **Strip or reject `store=true`** and `background=true` requests from ZDR agents, so behaviour does not depend on the provider's coercion.
5. **Keep state on your side.** Store conversation history and retrieval indexes in your own systems, under your own retention policy.
6. **Test it.** Send a request to a blocked endpoint with the agent's credential and confirm the gateway returns a policy denial and logs it.
7. **Re-check after framework upgrades.** New SDK versions can switch defaults to hosted state.

For how long your own logs should be kept, see [how long to retain AI agent logs](/blog/how-long-to-retain-ai-agent-logs). For limiting which models a project can call, see [restricting models in OpenAI API projects](/blog/restrict-models-in-openai-api-projects).

## Does the same apply to Azure OpenAI and other providers?

Yes, in principle. Microsoft's data-privacy page for models sold by Azure says Files, the Responses API, the Assistants API, stored completions, batch and fine-tuning store data in your Foundry resource until you delete it. That is separate from abuse monitoring. Each provider draws the line slightly differently, so build the endpoint allowlist per provider from its own documentation.

## How Opencontroller does this

Opencontroller auto-discovers models and API keys across AWS, Azure and Google Cloud, and records them in a governed registry next to the agents that use them. Its LLM Gateway puts control in the path — allow, route, redact or block — so an endpoint allowlist for ZDR agents is enforced on every request, with cost attributed per agent and per version. Policies are written as code. See the [capability reference](/capability-reference) and the [control plane overview](/blog/what-is-an-ai-agent-control-plane), or the [cross-platform governance guide](/blog/how-to-govern-ai-agents-across-platforms). To see it on your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does zero data retention cover the OpenAI Assistants API?**
No. OpenAI lists `/v1/assistants`, `/v1/threads` and `/v1/vector_stores` as not ZDR-eligible. Their objects are kept until you delete them.

**Is the Responses API ZDR-eligible?**
Yes, with limits. Under ZDR, `store` is treated as false. Background mode still keeps data on disk for about 10 minutes for polling.

**Can one OpenAI organisation mix ZDR and non-ZDR projects?**
Yes. The controls can be set per project, so verify the project an agent uses rather than assuming the organisation setting applies.

## Key facts

- OpenAI API abuse-monitoring logs are kept for up to 30 days by default; Zero Data Retention (ZDR) and Modified Abuse Monitoring need prior approval from OpenAI.
- Under ZDR, the store parameter on /v1/responses is always treated as false.
- OpenAI documents /v1/files, /v1/vector_stores, /v1/assistants, /v1/threads, /v1/conversations, /v1/batches, /v1/fine_tuning/jobs and /v1/evals as not ZDR-eligible.
- ZDR can be set per OpenAI organisation or per project.
- A ZDR agent must be blocked from stateful endpoints at the gateway, not just told not to call them.
- Opencontroller's LLM Gateway puts control in the path (allow, route, redact, block) and attributes cost per agent and version.
- Opencontroller starts from $250K per deployment per year (annual licence per deployment, not per seat or usage).

## Related

- [How do you restrict models and rate limits per OpenAI API project?](https://opencontroller.lyzr.com/blog/restrict-models-in-openai-api-projects.md)
- [How do you track OpenAI API costs by project with the Usage and Costs APIs?](https://opencontroller.lyzr.com/blog/track-openai-api-costs-by-project.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
