Claude Managed Agents Gets Built-In Memory: Cross-Session Learning for Production Agents

A Long-Standing Agent Problem, Now Solved

One of the most persistent limitations of production AI agents has been amnesia: every new session starts cold, with no memory of past mistakes, learned preferences, or domain context. Anthropic addressed this directly on April 23, 2026 with the public beta launch of built-in memory for Claude Managed Agents.

The premise is straightforward — agents that run repeatedly on the same workloads should get better over time, not restart with zero context. Memory on Claude Managed Agents allows agents to accumulate and retrieve knowledge across sessions: user preferences, project conventions, prior errors, and domain-specific patterns.

How It Works

Memory stores are mounted directly onto a filesystem, which means agents interact with them using the same bash and code execution tools they already employ for other tasks. There is no new API surface to learn — an agent that knows how to read and write files can read and write memories.

Developers attach one or more memory stores to an agent session via the Claude Console or the CLI. Up to 8 memory stores are supported per session. Stores can be scoped differently across agents: for example, an organization-wide store might be read-only for most agents while a user-specific store is read-write for a particular workflow.

Because memories are stored as files, they are fully portable. Developers can export memory stores, manage them via the API, and control exactly what agents retain — there is no black-box "neural" memory that cannot be inspected or edited.

Enterprise Controls Built In

Anthropic designed the memory layer for production deployments from the start. Every write to a memory store is tracked in a detailed audit log that records which agent and session produced the change. Teams can roll back a store to any prior state or redact specific content from history — a practical requirement for organizations that need to comply with data governance policies.

Multiple agents can be granted different access levels to shared memory stores. An organization might maintain a global store of project conventions accessible to all agents as read-only, while individual user stores allow each agent to accumulate personalized context.

Real-World Results

Four companies shared results from early deployments:

• Rakuten used memory to let long-running task agents learn from every session. The result: 97% fewer first-pass errors, 27% lower cost, and 34% lower latency.
• Wisedocs applied cross-session memory to a document verification pipeline, allowing agents to recognize and remember common issue patterns. Verification speed improved by 30%.
• Netflix used memory to retain context across sessions without requiring manual prompt engineering on each run.
• Ando eliminated the need for a custom memory infrastructure layer entirely by relying on Managed Agents's built-in memory.

Getting Started

Memory for Claude Managed Agents is available in public beta today. Developers can enable it via the Claude Console or the new CLI tool. Documentation is available at the Claude Platform docs under "Using agent memory."