Anthropic has announced a new feature for its Claude Platform that allows so-called Managed Agents to utilize built-in memory. The functionality is available in public beta and is designed for agents that need to retain and utilize information across multiple sessions.
The memory mechanism is designed as a separate layer intended to combine performance and flexibility. Instead of volatile context per session, agents can now store data and reuse it later. That data is saved as files, allowing developers to export and manage it via an API. This ensures that control over what an agent remembers remains with the user or developer.
According to Anthropic, this approach aligns with how agents are deployed in practice. Managed Agents run on infrastructure optimized for long-running tasks. The addition of memory is intended to ensure that these systems not only perform tasks but also learn from previous interactions and share that knowledge. Agents can simultaneously access the same memory storage without overwriting each other’s data.
Integration with existing tooling
The memory layer is linked to a file system, allowing agents to use the same resources as in regular tasks, such as bash commands and code execution. This should lead to better-structured storage of information and makes it possible to manage memory outside of runtime.
Storage can be configured with different permissions, such as read-only for organization-wide data and read-write permissions for user-specific data. All changes are logged, with each write action appearing as a session event in the Claude Console. This allows developers to track what an agent has learned and where that information comes from, as well as restore previous versions or delete content.
The memory files can be exported and managed via the API. In addition to the console, the platform also offers a command-line interface for deploying and testing agents with memory.
This feature responds to a broader trend in which AI agents increasingly exhibit persistent behavior and build upon previous interactions. This is relevant for applications such as document processing, customer service, and internal tooling.
With the introduction of built-in memory, Anthropic is taking a step toward more autonomous agents. At the same time, the question remains how these types of systems relate to requirements regarding transparency, control, and data management.