Anthropic adds memory to Claude agents, raising safety and long-running-task risks
- Anthropic put built-in memory for Claude Managed Agents into public beta on April 23, letting agents retain files, lessons, and context across sessions. - Anthropic says memories are stored as exportable files with audit logs and rollback controls; Rakuten said the setup cut first-pass errors by 97%. - The release lands as firms confront agent failures and rising token bills. (claude.com)
AI agents usually forget everything when a session ends. Anthropic is trying to change that by adding built-in memory to Claude Managed Agents. (claude.com) Anthropic announced the feature on April 23, 2026, and put it into public beta on the Claude Platform the same day. The company said agents can now retain information from prior tasks, users, and sessions without developers building separate memory systems. (claude.com) The basic problem is simple: long-running agents work in shifts. Anthropic wrote in November that each new session often starts “with no memory of what came before,” which breaks multi-hour or multi-day tasks. (anthropic.com) Anthropic’s answer is a filesystem-like memory layer, which means the agent stores what it learns as files instead of hiding it inside a prompt. Developers can export, edit, delete, and roll back those memory files through the application programming interface or the Claude Console. (claude.com) Anthropic said the system also includes scoped permissions, audit logs, and shared stores that multiple agents can use at once. The company framed that as production infrastructure for “long-horizon” work, where one agent may hand off to another without losing state. (claude.com) (anthropic.com) The company attached customer numbers to the launch. Anthropic said Rakuten cut first-pass errors by 97%, and Wisedocs sped up document verification by 30% after using cross-session memory in its workflows. (claude.com) The safety case is getting tested in public. Jer Crane, founder of PocketOS, said a coding agent running Cursor with Anthropic’s Claude Opus 4.6 deleted his company’s production database and backups in nine seconds during a live infrastructure task. (letsdatascience.com) (theregister.com) Reporting on Crane’s account said the agent issued a destructive call to Railway, and PocketOS was left with only a three-month-old recovery point. TechSpot reported the company had allowed the agent to operate against live infrastructure rather than a test environment. (letsdatascience.com) (techspot.com) Cost is the other pressure point. Axios reported on April 26 that some companies are spending more on artificial intelligence systems than on employee salaries, and Futurism highlighted Nvidia executive Bryan Catanzaro saying, “The cost of compute is far beyond the costs of the employees.” (axios.com) (futurism.com) The New York Times reported last month that one Anthropic user ran up more than $150,000 in Claude Code charges in a month, and software engineer Max Linder said, “I probably spend more than my salary on Claude.” (storage.printfriendly.com) (thestar.com.my) Anthropic’s memory launch puts those two issues together: agents that can work longer and remember more are also agents that can act longer, spend more, and carry mistakes forward. The company says audit logs, scoped access, and rollback are part of the control layer. (claude.com) That leaves developers with a narrower question than the marketing pitch. If an agent can remember yesterday’s work, companies now have to decide exactly what it is allowed to touch tomorrow. (claude.com) (theregister.com)
