Anthropic · 2026-04-23 · notable
Anthropic Managed Agents Memory: Persistent Stores for Claude Agent Sessions
Claude Managed Agents sessions can now attach persistent memory stores: workspace-scoped document collections that survive across sessions. Agents read and write them as a /mnt/memory/ directory. 8 stores per session, 100KB per memory, 250k immutable version history with point-in-time recovery. Public beta as of April 23.
Claude Managed Agent sessions now persist user preferences, project context, and domain knowledge across sessions — the agent mounts them as a filesystem directory.
Key specs
| Memory stores per session | 8 |
|---|---|
| Max size per memory | 100 KB |
| Memories per store | 2,000 |
| Storage per store | 100 MB |
| Version history per store | 250,000 |
What is it?
Anthropic released memory stores for Claude Managed Agents in public beta on April 23, 2026. A memory store is a workspace-scoped collection of text documents that survive after a session ends. When attached to a new session, the agent reads and writes it as a directory at /mnt/memory/ using the same file tools it uses for everything else. Up to 8 stores can be attached per session, with access modes of read_write or read_only.
How does it work?
You create a store with a name and description via the Admin API (or the ant CLI). Attach it to a session in the resources[] array when creating the session. The agent treats each mounted store as a regular filesystem path — it can create, read, update, delete, and grep files in the store using the standard agent toolset. Every write creates an immutable memver_... version with 30-day retention, giving a full audit trail with SHA-256 precondition support for safe concurrent writes. Memory content is also manageable directly via the API — useful for seeding stores before any agent runs, reviewing what an agent wrote, or scrubbing PII via the redact endpoint.
Why does it matter?
Without persistent memory, every Managed Agents session starts cold — the agent cannot remember user preferences, project conventions, or findings from prior sessions. Memory stores close this gap for production agent deployments. The versioning system makes it safe to give agents write access: you can inspect every mutation and roll back. The read_only mode is a practical safeguard for shared reference material against prompt injection attacks.
Who is it for?
Developers building production agents with the Anthropic Managed Agents API
Try it
pip install anthropic && see platform.claude.com/docs/en/managed-agents/memory