Anthropic gives Claude agents persistent memory

Anthropic said on April 23 that Claude Managed Agents now have built-in memory in public beta, so agents can carry context from one session to the next. (claude.com) In plain terms, agent memory is a saved working notebook: instead of starting blank every time, the system can keep facts, corrections, and prior decisions. Anthropic said its version stores memory as files mounted on a filesystem the agent can read and update with the same bash and code tools it already uses. (claude.com) Anthropic said developers can export those memory files, manage them through the application programming interface, and control what an agent keeps. The company also said memory changes come with audit logs, version rollback, redaction tools, and per-store permissions. (claude.com) That matters for the kind of work Anthropic is targeting with Managed Agents, its hosted service for long-horizon tasks launched in public beta on April 8. Anthropic said the service is built around stable interfaces for sessions, harnesses, and sandboxes so the underlying implementation can change without breaking customer workflows. (anthropic.com) Before this update, Anthropic said one of the main production gaps was that agents lost what they learned when a session ended. The new memory layer is meant to let multiple agents share what they have learned across sessions without maintaining separate custom infrastructure. (claude.com) (anthropic.com) Anthropic tied that design to enterprise controls. The company said stores can be shared across agents with different read and write scopes, and multiple agents can work against the same store without overwriting each other. (claude.com) The company also pointed to customer results. Anthropic said Rakuten’s long-running agents cut first-pass errors by 97%, while Wisedocs used cross-session memory in document verification and sped verification by 30%. (claude.com) The memory launch builds on Anthropic’s earlier Agent Skills work, published in October 2025, which packages instructions, scripts, and resources into folders a model can load when needed. Anthropic describes a skill as a directory with a `SKILL.md` file and related resources, like an onboarding guide for a new hire. (anthropic.com) Put together, Skills give an agent reusable procedures, while memory gives it a record of what happened last time. Anthropic’s pitch is that developers can combine the two to build agents for repeatable work instead of one-off demos. (anthropic.com) (claude.com) Anthropic’s next test is whether those controls and customer gains hold up beyond beta. For now, the company is selling a more durable kind of agent: one that does not wake up empty-handed each time it starts. (claude.com)