What is Vertex AI Memory Bank and how does it work?

Short answer

Memory Bank is Google Cloud's managed long-term memory service for AI agents, launched as part of Vertex AI Agent Engine and now documented as Agent Platform Memory Bank in Gemini Enterprise Agent Platform. It uses Gemini models to extract facts and preferences from conversations, merges them with existing memories per scope, such as a user ID, and lets an agent retrieve them later by similarity search.

First, the name

If you search for “Vertex AI Agent Engine Memory Bank” you will find two sets of names. Google launched Memory Bank in public preview in July 2025 as part of Vertex AI Agent Engine. Google’s current documentation calls the platform Gemini Enterprise Agent Platform and the service Agent Platform Memory Bank, and the ADK docs call the runtime Agent Runtime. The code still uses the old names, such as agent_engine_id and the agentengine:// URI. It is the same service.

What it does

Memory Bank gives an agent long-term memory across conversations. Google’s launch post describes the problem it solves: without memory, agents “treat each interaction as the first”, and pasting whole past conversations into the context window is expensive, slow and lowers quality.

The flow has four parts:

  1. Sessions hold the conversation. Each chat between a user and an agent is stored as a session with its events.
  2. Gemini extracts memories. Memory Bank reads the session history and pulls out key facts, preferences and context, such as “I prefer aisle seats on flights.” According to the launch post this happens asynchronously in the background.
  3. New facts are merged with old ones. When new information arrives, Memory Bank uses Gemini to consolidate it with existing memories, “resolving contradictions and keeping the memories up to date.” Memories are organized by a scope you define, such as a user ID.
  4. The agent retrieves them later. Either all facts for a scope, or a similarity search with embeddings for the memories most relevant to the current topic.

Google says the extraction method is based on Google Research work accepted at ACL 2025.

How you use it with ADK

In Google’s Agent Development Kit, memory is a pluggable service:

Option What it is When to use it
InMemoryMemoryService Keeps session information in the application’s memory with basic keyword matching; lost on restart Prototyping and tests
VertexAiMemoryBankService Connects to Memory Bank: extraction, consolidation and semantic search Persistent memory in production
load_memory tool The agent decides itself when to look something up When memory is only sometimes relevant
PreloadMemoryTool Retrieves memory automatically at the start of every turn When memory should always be considered

You add memory with add_session_to_memory after a conversation, or with add_memory for explicit entries. Setup needs a Google Cloud project with the Agent Platform API enabled, an Agent Runtime instance and its ID, and local authentication. For other frameworks, Google’s launch post points to calling the Memory Bank API directly.

The common misunderstanding

Memory Bank is not a place to store your company’s documents. It is built for what users say in conversations: preferences and facts about a person, extracted by a model and kept per scope. Google’s documentation summarizes it as a service to “store long-term user preferences and facts.” If you want an agent to answer from your policies, contracts or wiki, that is a retrieval question, not a memory question. The difference is explained in Is RAG the same as AI memory?

A second point: because a model decides what becomes a memory, a false statement in a conversation can end up stored as fact. That risk is covered in What is memory poisoning in AI agents?

What this means for a team

Memory Bank is a building block for teams that build their own agents on Google Cloud. Memories are kept per scope, usually per user, so the default is personal memory, not shared team knowledge. You can choose a broader scope, but then you decide yourself who may read and write what; the documentation points to IAM Conditions for access control. It also only reaches agents that call it. What your team tells ChatGPT, Claude or Copilot stays in those tools. For the wider picture, see What are the types of memory in AI agents?

Where Bombos fits

If you build agents on Google Cloud and need per-user memory, Memory Bank is a sensible managed option. Bombos solves a different layer: one shared memory for all the AI tools a team already uses, including Claude, ChatGPT, Copilot, Cursor and your own agents, with team knowledge, learned rules about how the team works, and what already sits in mail, documents and repos. You can see which knowledge an answer used, and it runs in our cloud or on your own server.

Related questions

Do I need to deploy my agent on Google Cloud to use Memory Bank?

No. The ADK documentation says you do not need to deploy your agent to Agent Runtime to use Memory Bank; you do need a Google Cloud project with the API enabled and an Agent Runtime instance whose ID you pass to the memory service.

Does Memory Bank only work with Google's ADK?

No. ADK has a built-in VertexAiMemoryBankService, but Google's launch post says you can also call Memory Bank from other frameworks such as LangGraph and CrewAI, and there are sample notebooks for both.

What is the difference between Sessions and Memory Bank?

Sessions store the individual interactions of one conversation. Memory Bank draws long-term memories out of those sessions so they are available in later conversations.

Is Memory Bank data kept in my region?

Google's documentation lists data residency at rest, CMEK, VPC Service Controls and HIPAA support for Memory Bank, and says its machine learning processing happens in the region or multi-region of the model endpoint it uses.

Sources

  1. Memory Bank overview (Google Cloud documentation)
  2. Vertex AI Memory Bank in public preview (Google Cloud blog, July 2025)
  3. Memory (Agent Development Kit documentation)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.