Anthropic opened a private beta for Agent Memory, a managed service that gives Claude agents a persistent memory across sessions. Until now every agent conversation started from zero: whatever the agent learned about a codebase, a customer or a workflow disappeared the moment the session ended.
What changed
Agent Memory extracts the key information from agent conversations and surfaces it again when it becomes relevant, instead of replaying entire transcripts. The practical effect is on the context window — the agent keeps what matters without carrying the full history, so context stays lean and cost per run stays predictable.
Memory is built as a filesystem-based layer: entries are stored as files that can be exported, managed through the API and scoped with permissions. Every change is logged, with an audit trail per session and per agent, so teams can roll back, redact or remove what an agent has retained.
Availability
The service started as a private beta and moved to public beta on the Claude Platform on April 23, 2026, for Claude Managed Agents. Early adopters named by Anthropic include Netflix, Rakuten, Wisedocs and Ando, using it to cut repeated setup work and reduce errors in long-running workflows.
Alongside the announcement Anthropic opened a $100K developer hackathon built around Claude Opus 4.7.
Why it matters
Persistent memory is what separates a chatbot from an agent that gets better at its job. Without it, every improvement has to be pushed manually into prompts; with it, an agent accumulates knowledge on its own and the operator stops re-explaining the same context. For teams running agents on support queues, code review or document processing, this removes the largest source of repetitive prompt maintenance.
The audit trail matters just as much as the memory itself: anything an agent remembers about customers or internal systems has to be inspectable and deletable before an enterprise will put it into production.