---
title: How do you get zero data retention for AI agents on Google Vertex AI?
url: https://opencontroller.lyzr.com/blog/vertex-ai-zero-data-retention-for-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you get zero data retention for AI agents on Google Vertex AI?

**To get zero data retention on Vertex AI, disable project caching, keep request-response logging and Live API session resumption off, set `store=false` on the Interactions API, and avoid features whose storage cannot be disabled: Grounding with Google Search, Grounding with Google Maps, Deep Research and the Sandbox Snapshot API. Abuse-monitoring prompt logging may still apply, and OpenAI partner models need a separate ZDR approval.**

Google's documentation now calls the product Gemini Enterprise Agent Platform, though API names still use `aiplatform`. The settings below apply either way. For the wider pattern, see [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms).

## What does Google retain by default?

Google states it will not use your data to train or fine-tune models without your permission. Retention is a separate question. Several features keep data for short periods by default or by design.

| Feature | Default | Retention | Can you turn it off? |
|---|---|---|---|
| In-memory caching (supported Gemini models) | On | Up to 24 hours, memory only | Yes |
| Durable caching | On | Within 24 hours, encrypted at rest | Yes |
| Request-response logging to BigQuery | Off | Your BigQuery retention | Keep it off |
| Live API session resumption | Off | Up to 24 hours | Keep it off |
| Interactions API `store` | `true` | Stored | Set `store=false` |
| Grounding with Google Search | Per request | Up to 3 days | No |
| Grounding with Google Maps | Per request | 30 days | No |
| Deep Research agent | Per use | 7 days | No |

Caching is scoped to your project, never served to another project, and processed in the serving region.

## How do you switch off caching?

A user with `roles/aiplatform.admin` sends a PATCH to the project's `cacheConfig` endpoint with `"disableCache": true`. That turns off both in-memory and durable caching.

If you want the speed of in-memory caching but no data at rest, set `retentionType` to `EPHEMERAL` instead. That keeps in-memory caching and disables durable caching. Disabling durable caching across a whole organization requires your Google account team.

Note that durable caching does not support customer-managed encryption keys (CMEK).

## Which agent features break zero data retention?

Agents often switch these on without anyone noticing:

1. **Search grounding.** An agent that "checks the web" with Google Search grounding stores derived queries for up to 3 days. Google recommends Web Grounding for Enterprise when you need ZDR.
2. **Maps grounding.** Prompts, context and output are stored for 30 days.
3. **Deep Research.** Session data is kept for 7 days.
4. **Live API session resumption.** Keeps text, audio, video and outputs for up to 24 hours.
5. **Interactions API with default `store`.** Must be set to `false`.
6. **Request-response logging.** Useful for audit, but it writes payloads to BigQuery. See [logging Vertex AI prompts to BigQuery](/blog/log-vertex-ai-prompts-and-responses-to-bigquery) for the trade-off.

## What about partner models?

OpenAI flagship models on the platform have their own rules:

- ZDR is **off by default** and only available on approved projects.
- Without it, prompts, file inputs and responses are logged for up to 30 days in the `us` multi-region for safety and abuse monitoring.
- Existing OpenAI API approvals do not carry over, and `store: false` alone does not enable ZDR.
- Approval takes a request to your account team with project IDs. Configuration takes up to 24 hours to propagate.
- A "Zero Data Retention active" banner on the model card confirms it.

## What ZDR checklist should an agent team follow?

1. Confirm whether your terms include abuse-monitoring prompt logging for Google models, and record the answer.
2. Disable caching (or set `EPHEMERAL`) on every project agents use.
3. Keep request-response logging off, or document why audit needs it.
4. Set `store=false` on Interactions API calls.
5. Block Search and Maps grounding for ZDR agents, or switch to Web Grounding for Enterprise.
6. Block Deep Research and the Sandbox Snapshot API for those agents.
7. Request partner model ZDR per project where needed.
8. Pin regions. See [the global endpoint and data residency guide](/blog/vertex-ai-global-endpoint-and-data-residency).
9. Re-check after every new feature or model is enabled.

Across providers, the same pattern repeats: a few features sit outside ZDR. The [cross-provider exclusions guide](/blog/llm-api-features-excluded-from-zero-data-retention) lists the common ones.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Google Cloud, AWS and Azure, and has runtime adapters for Vertex AI, Bedrock AgentCore and Microsoft Foundry. Posture detects risky permission and configuration settings. Policy as code can encode rules such as "ZDR agents may not use grounding tools", and the LLM Gateway puts control in the path to allow, route, redact or block. [Book a demo](https://www.lyzr.ai/book-demo/) to see it on your projects.

## Frequently asked questions

**Does Google train on Vertex AI prompts?**
Google states it will not use your data to train or fine-tune models without your prior permission or instruction.

**Can we keep in-memory caching and still avoid data at rest?**
Yes. Set `retentionType` to `EPHEMERAL`, which keeps in-memory caching and disables durable caching.

**Does ZDR remove all logging?**
No. Operational and billing metadata is still recorded, and abuse-monitoring prompt logging may apply to Google models depending on your terms.

## Key facts

Google does not train on Vertex AI (Gemini Enterprise Agent Platform) customer data without permission. For zero data retention with Google models: disable caching via the project cacheConfig (disableCache true, requires roles/aiplatform.admin); in-memory and durable caches are otherwise deleted within 24 hours. Do not enable request-response logging to BigQuery or Live API session resumption; set store=false on the Interactions API; avoid Grounding with Google Search (queries stored up to 3 days, cannot be disabled), Grounding with Google Maps (30 days), Deep Research (7 days) and the Sandbox Snapshot API. Prompt logging for abuse monitoring may still apply. OpenAI partner models need a separate ZDR approval; without it prompts and responses are logged for up to 30 days. Opencontroller provides runtime adapters for Vertex AI. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [Does the Vertex AI global endpoint break data residency for agents?](https://opencontroller.lyzr.com/blog/vertex-ai-global-endpoint-and-data-residency.md)
- [Which LLM API features are excluded from zero data retention?](https://opencontroller.lyzr.com/blog/llm-api-features-excluded-from-zero-data-retention.md)
- [How do you log Vertex AI prompts and responses to BigQuery for audit?](https://opencontroller.lyzr.com/blog/log-vertex-ai-prompts-and-responses-to-bigquery.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
