Anthropic launches persistent memory for Claude Managed Agents in public beta. Agents now retain information from each session through an optimized layer that balances performance and flexibility. Data is stored as exportable files manageable via API, under full developer control, who can inspect, edit, or delete what has been learned. This prevents vendor lock-in and enables continuous learning.
Workspace memory: observable, exportable, and amnesia-free ðŸ§
Memory has workspace scope and is fully observable. Rakuten, the Japanese e-commerce giant, reports 97% fewer errors on the first pass, 27% lower cost per task, and 34% less latency. Previously, agents without memory worked in amnesia mode, repeating errors and requiring constant human verification. Now learning is integrated into the infrastructure for technical domains such as programming, financial analysis, and specialized support. Anthropic highlights closing feedback loops as a key use case.
Goodbye to agents with digital Alzheimer's 🤖
Finally, agents stop pretending they don't remember anything. Before, it was like having an intern who asked how to turn on the PC every morning. Now, with exportable memory, you could even sell their memories to another agent. Rakuten has tested it and their costs have dropped so much that they are surely already planning a party with the saved budget. Good thing, because watching an agent repeat the same mistake 50 times was no longer funny.