Anthropic Unveils Memory

Anthropic introduced Memory as its next-generation agent primitive, adding it to the managed agents API in public beta.

Memory is designed to let agents store and update task success criteria, recurring mistakes, effective strategies, and codebase/file/asset information in the form of a file system. Agents work directly with memory files using bash and GP tools, with the goal of supporting long-term learning and context management in large-scale multi-agent environments.

The key design elements are as follows.

• Permission scoping: organizational knowledge is read-only, while working memory is read-write
• Optimistic concurrency: content hashes prevent conflicts from simultaneous updates
• Version history / attribution: tracks who changed what, when, and in which session
• Standalone API: enables PII checks, cleanup, and replication to external systems even outside managed agents

Anthropic also released Dreaming as a research preview. This process analyzes recent agent sessions and transcripts to identify patterns and mistakes, automatically generating organized memory. In early testing, the company reported a 6x increase in task completion rate in a specific scenario on Harvey's legal benchmark.