Anthropic adds persistent agent memory
- Anthropic said April 23 that persistent memory is now in public beta for Claude Managed Agents, letting agents keep and reuse information across sessions. - Anthropic stores those memories as files developers can export and edit; Rakuten said the setup cut first-pass errors by 97%. - The release pushes agent design toward stateful, auditable workflows with shared stores and rollback controls. (claude.com)
Anthropic put persistent memory into Claude Managed Agents on April 23, moving its hosted agents beyond one-session chats and into cross-session work. (claude.com) Managed Agents already ran long-horizon tasks on Anthropic’s infrastructure. The new memory layer lets those agents keep user preferences, project conventions, prior mistakes, and domain context after a session ends. (anthropic.com) (platform.claude.com) Anthropic says the feature is in public beta and available through the Claude Platform. Managed Agents API requests use the `managed-agents-2026-04-01` beta header, which Anthropic says its software development kits set automatically. (claude.com) (platform.claude.com) The basic idea is simple: a normal agent session starts with a blank slate, like a worker clocking in with no notes from yesterday. A memory store gives that agent a saved folder it can reopen in later sessions. (platform.claude.com) Anthropic built that folder as a filesystem mount inside the agent’s container, so Claude reads and writes memory with the same file tools it already uses for code and documents. Anthropic says that design helped its latest models save more organized memories and decide more selectively what to keep. (platform.claude.com) (claude.com) The company is pitching control as much as convenience. Memories can be exported, edited through the application programming interface or console, shared across agents with read-only or read-write permissions, and rolled back through immutable version history. (claude.com) (platform.claude.com) Anthropic also says every change is logged with the agent and session that made it. That gives developers an audit trail for what an agent learned, when it learned it, and where that information came from. (claude.com) (platform.claude.com) The company used customer examples to show the pitch. Anthropic said Rakuten’s long-running agents cut first-pass errors by 97%, while Wisedocs sped document verification by 30% after using cross-session memory to remember recurring issues. (claude.com) Those examples also show what changes when agents stop forgetting. Instead of stuffing every rule into a prompt each time, developers can keep smaller, scoped files that persist across sessions; Anthropic caps each individual memory file at 100 kilobytes, or about 25,000 tokens. (platform.claude.com) Anthropic launched Claude Managed Agents in public beta earlier in April. Adding memory less than three weeks later makes retention, permissions, and audit logs part of the core product, not extra plumbing developers bolt on later. (anthropic.com) (claude.com)
