Anthropic released persistent memory for Claude Managed Agents in public beta, enabling agents to learn from past sessions and share what they’ve learned with other agents. The feature, announced on the Claude blog, stores memories as files on a mounted filesystem, giving developers programmatic control over what agents retain.
How It Works
Memory mounts directly onto the Managed Agents filesystem, building on the same bash and code execution tools that agents already use, Anthropic wrote. Multiple agents can work against the same memory store concurrently without overwriting each other.
The system supports scoped permissions: organization-wide stores can be set as read-only while per-user stores allow reads and writes. All changes are tracked through a detailed audit log that records which agent and session a memory originated from, per SD Times. Developers can roll back to earlier versions or redact specific content from history. Updates surface in the Claude Console as session events.
Memories can be exported and managed independently via API, according to Anthropic. The company emphasized that the feature is “optimized against internal benchmarks for long-running agents that improve across sessions.”
Production Deployments
Three companies are already using the feature in production, per the Anthropic announcement:
Netflix agents carry context across sessions, including insights that required multiple turns to uncover and corrections made by humans mid-conversation, replacing manual prompt and skill updates.
Rakuten’s task-based agents use memory to learn from every session and avoid repeating past mistakes. Yusuke Kaji, General Manager of AI for Business at Rakuten, said the feature delivers “97% fewer first-pass errors at 27% lower cost and 34% lower latency,” with continuous learning remaining “workspace-scoped and observable.”
Wisedocs built its document verification pipeline on Managed Agents, using cross-session memory to identify and remember recurring document issues, speeding up verification by 30%.
Competitive Context
The release comes weeks after Anthropic launched Managed Agents in early April, as 9to5Mac noted. Memory addresses a core limitation of stateless agent architectures: without persistence, every session starts from zero. For enterprise teams running agents across hundreds or thousands of daily interactions, that reset destroys accumulated knowledge about edge cases, user preferences, and organizational patterns.
OpenAI’s workspace agents, announced April 22, include their own memory capabilities. Google’s Gemini Enterprise Agent Platform, unveiled the same week at Cloud Next, also features persistent state management. The rapid convergence signals that statefulness is becoming a baseline expectation for production agent infrastructure, not a differentiator.