---
title: How do you govern Vertex AI Agent Engine Memory Bank?
url: https://opencontroller.lyzr.com/blog/vertex-ai-memory-bank-governance
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you govern Vertex AI Agent Engine Memory Bank?

**Govern Memory Bank through four settings: a scope on every memory (who it belongs to), TTLs on memories and revisions (how long it lives), purge and delete (how it leaves), and IAM Conditions plus CMEK and VPC Service Controls (who can reach it). The trap is revisions: they keep snapshots, including extracted source content, for 365 days by default and survive deletion for up to 48 hours.**

Google now documents Memory Bank under Gemini Enterprise Agent Platform; resources still sit under `reasoningEngines`, the API used by Vertex AI Agent Engine.

## What does Memory Bank store?

Memory Bank stores long-term facts and preferences that an agent learns about a user. Memories are created directly, or generated by a model from conversation events. During generation, Memory Bank extracts candidate facts and consolidates them with existing memories in the same scope, so it can create, update or delete memories automatically.

Each memory has:

| Field | What it means | Governance impact |
|---|---|---|
| `scope` | Required, immutable key-value pairs, such as a user ID | Memories are isolated within their scope; retrieval matches it exactly |
| `fact` | The remembered content | May contain personal data |
| `topics` | Managed or custom topic labels | Limit what can be generated or retrieved |
| `ttl` / `expireTime` | When the memory expires | Your retention control |
| Revisions | Snapshots at each change | A second copy of the data with its own TTL |

## How do you control retention?

1. **Set a TTL on memories.** You can pass a `ttl` per memory, and the Memory Bank configuration supports default and per-action TTLs at instance level.
2. **Set a revision TTL.** Revisions are on by default with a 365-day TTL. Set `revision_ttl` to match your retention policy, or disable revisions with `disable_memory_revisions` where you do not need rollback.
3. **Know what revisions hold.** For memories changed by generation, a revision also stores the `extracted_memories` taken from the source before consolidation. That is closer to raw conversation content than the final fact.
4. **Use topics to stop collection.** If a category of data should never be remembered, don't enable a topic for it.

## How do you delete a user's memories on request?

Plan for data subject requests before launch:

1. **Scope by a stable user key.** If memories are scoped by user ID, one filter finds everything for that person.
2. **Purge by filter.** The `purge` method takes a required AIP-160 `filter`. With `force` set to false it only validates; set it to true to delete.
3. **Account for the recovery window.** Deleting a memory writes a final revision with an empty fact. Prior revisions remain accessible for up to 48 hours, and you can roll back within that window. Your erasure record should note when that window closes.
4. **Check other stores.** Agent Engine Sessions, logs and traces may hold the same conversation. Memory Bank deletion does not reach them.
5. **Record the action.** Log the request, filter, operation result and timestamp as evidence.

## Who can read and write memories?

- **IAM Conditions** can restrict Memory Bank access, so a service or user is limited to particular memories.
- **CMEK** is supported. For deployed agents it works only on single-region endpoints, not multi-region. Standalone instances support regional and `us` or `eu` multi-region endpoints, not `global`.
- **VPC Service Controls** are supported.
- **Data residency at rest** is supported; model processing for generation happens in the region of the model endpoint Memory Bank uses.
- Google also lists HIPAA, Access Transparency and Access Approval support.

## What risks remain?

- **Memory poisoning.** Generated memories come from conversation content, which can include injected instructions. A false "fact" is then retrieved in every future session for that scope. See [how to prevent memory poisoning in AI agents](/blog/prevent-memory-poisoning-in-ai-agents).
- **Wrong scope.** An application that scopes by tenant instead of user shares one user's facts with colleagues.
- **Silent accumulation.** Without TTLs, memories never expire.

Use rollback to recover from a bad generation and review high-impact topics regularly. For general patterns across platforms, see [how to govern AI agent memory stores](/blog/govern-ai-agent-memory-stores).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Google Cloud, AWS and Azure, and has a runtime adapter for Vertex AI. Agents land in a governed registry with owners, versions and tools, and Posture flags risky permissions and configurations. Policy as code with built-in mappings to NIST, ISO 42001, OWASP and MITRE ATLAS helps turn retention rules into checks. See [how to govern Vertex AI Agent Engine agents](/blog/govern-google-vertex-ai-agent-engine-agents), [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms) and [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane). [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How long do Memory Bank revisions last?**
Revisions are enabled by default with a 365-day TTL. You can change it with `revision_ttl` or disable revisions instance-wide or per request.

**Does deleting a memory remove it immediately?**
The memory returns 404 at once, but its revisions stay accessible for up to 48 hours, during which you can roll back.

**Can Memory Bank use customer-managed encryption keys?**
Yes, with limits. For deployed agents, CMEK works only on single-region endpoints, not multi-regional ones.

## Key facts

Google Memory Bank (Vertex AI Agent Engine, now documented under Gemini Enterprise Agent Platform) isolates memories by a required, immutable scope. Memories support ttl/expireTime; revisions are on by default with a 365-day TTL and store extracted_memories for generated changes. Deleted memories keep revisions for up to 48 hours. purge needs an AIP-160 filter and force=true to delete. IAM Conditions, VPC-SC and CMEK (single-region for deployed agents) are supported. Opencontroller has a Vertex AI runtime adapter. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern what AI agents remember?](https://opencontroller.lyzr.com/blog/govern-ai-agent-memory-stores.md)
- [How do you stop memory poisoning in AI agents?](https://opencontroller.lyzr.com/blog/prevent-memory-poisoning-in-ai-agents.md)
- [How do you govern agents deployed on Vertex AI Agent Engine?](https://opencontroller.lyzr.com/blog/govern-google-vertex-ai-agent-engine-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
