TLDR: Google has announced the public preview of ‘Memory Bank’ within its Vertex AI Agent Engine, a new feature designed to provide AI agents with persistent, long-term memory. This innovation addresses a critical limitation of conversational AI – the inability to recall past interactions and user preferences – by allowing agents to retain context across sessions, leading to more personalized and continuous user experiences. The service leverages Gemini models for intelligent memory extraction and management, and integrates with Google’s Agent Development Kit (ADK) as well as popular frameworks like LangGraph and CrewAI.
Google has taken a significant leap in the evolution of artificial intelligence with the public preview announcement of ‘Memory Bank,’ a groundbreaking new capability within its Vertex AI Agent Engine. Unveiled on July 8, 2025, this managed service is poised to revolutionize how AI agents interact with users by providing them with persistent, long-term memory, effectively solving one of the most fundamental challenges in conversational AI: the lack of contextual recall.
Historically, AI agents have been likened to ‘digital goldfish,’ treating each interaction as a fresh start. This limitation often leads to repetitive questioning, a failure to remember user preferences, and a generally disjointed user experience. As Google explains, ‘without memory, agents treat each interaction as the first, asking repetitive questions and failing to recall user preferences.’ The conventional workaround, leveraging an LLM’s context window, proved to be expensive, inefficient, and prone to issues like ‘lost in the middle’ and ‘context rot,’ where output quality declines as context becomes diluted with irrelevant details.
Memory Bank directly tackles these issues. It functions by intelligently extracting key facts, preferences, and contextual information from a user’s conversation history, which is stored in Agent Engine Sessions. This extraction process is powered by Google’s advanced Gemini models and occurs asynchronously in the background, ensuring that the agent’s responsiveness is not compromised. The extracted memories are then stored persistently, organized by user ID, and can be updated and consolidated over time, even resolving contradictions to maintain an accurate and relevant memory profile.
This new capability offers a multitude of benefits for both developers and end-users. For users, it translates into highly personalized interactions, where agents remember past choices and preferences, and seamless session continuity, allowing conversations to resume smoothly even after days or weeks. The improved contextual awareness eliminates the need for users to repeat themselves, significantly enhancing the overall experience.
For developers, Memory Bank eliminates the ‘undifferentiated work’ of building and managing their own memory infrastructure, such as vector databases and retrieval logic. It provides a scalable, managed solution that is far more efficient than constantly populating large context windows. The service also boasts seamless integration, supporting Google’s Agent Development Kit (ADK) natively, as well as popular external frameworks like LangGraph and CrewAI.
Kimberly Milam, Software Engineer at Vertex AI, and George Lee, Product Manager for Cloud AI Research, highlighted the practical applications of Memory Bank. Examples include a beauty advisor agent that can adapt to a user’s evolving skincare routine, or an agent remembering specific preferences like ‘I prefer aisle seats’ or ‘my preferred temperature is 71 degrees.’ This intelligent recall is grounded in novel research from Google Research, accepted by ACL 2025, which introduces a topic-based approach to how agents learn and recall information.
Also Read:
- Google AI Releases Open-Source MCP Toolbox to Enhance Secure Database Access for AI Agents
- Google DeepMind Unveils Gemini Robotics On-Device: Empowering Robots with Local AI Capabilities
The public preview of Vertex AI Memory Bank marks a pivotal moment in the industry’s shift towards persistent context in AI applications, promising a future where AI agents are not just intelligent, but also truly remember and understand their users over time.
