← Back to feed

Anthropic has shipped persistent memory for Claude Managed Agents in public beta, giving developers a first-class mechanism for agents that improve across sessions rather than starting cold every time.

The feature works through memory stores: workspace-scoped collections of text files that mount inside the agent’s container at /mnt/memory/. The agent reads and writes them using the same file tools it uses for the rest of the filesystem. Writes persist across sessions. Up to 8 stores can be attached per session; individual memory files are capped at 100KB each.

Memory stores are created via API, attached when a session starts, and configured with read_write or read_only access. Once attached, the system automatically adds a prompt note describing each mount location so the agent knows where to look without explicit instruction. Stores can be shared across sessions — a pattern that enables fleets of agents to propagate lessons learned by one instance to all others.

What Early Adopters Reported

Four companies joined the early access program:

• Rakuten: 97% fewer first-pass errors, 27% cost reduction, 34% lower latency. Agents distill lessons from each session and apply them in subsequent runs rather than re-deriving context from scratch.
• Netflix: Agents carry context across sessions and apply mid-conversation human corrections without requiring manual system prompt updates.
• Wisedocs: Document verification pipeline runs 30% faster by storing domain-specific verification patterns across sessions.
• Ando: Eliminated the need for custom memory infrastructure built on top of the API.

How the Memory Layer Works

Memory stores are versioned. Every write creates an immutable version attributed to the session, with 30-day retention and full audit trail via the version endpoints. Content can be redacted from history when needed.

The store mounts as a directory, so the agent can organize memory as many small focused files rather than a single growing context dump. Anthropic describes this as enabling just-in-time context retrieval: the agent pulls only what it needs into the active window rather than loading everything upfront.

A security caveat: read_write stores are vulnerable to prompt injection. If the agent processes untrusted input — user prompts, fetched web content, third-party tool output — a successful injection can write malicious content into memory, which then surfaces as trusted context in every subsequent session. Anthropic recommends read_only for reference material or any store the agent does not need to modify.

The memory tool qualifies for Zero Data Retention. When ZDR is active, data is not stored after the API response is returned.

Claude Opus 4.7 Integration

Anthropic says Opus 4.7 is specifically optimized for filesystem-based memory operations. The model is more selective about what to retain and generates more organized, structured memory files compared to earlier versions. For long-running software projects, the pattern Anthropic recommends starts with an initializer session that writes a progress log, feature checklist, and startup reference before any substantive work begins. Subsequent sessions open by reading those artifacts, recovering full project state in seconds without re-exploring the codebase.

The beta header required for access is managed-agents-2026-04-01. During beta, session infrastructure storage and session-minutes are free; usage-based pricing is expected when the feature exits beta.

Key Numbers

• Rakuten first-pass error reduction: 97%
• Rakuten cost reduction: 27%
• Rakuten latency reduction: 34%
• Wisedocs verification speed improvement: 30%
• Memory stores per session: up to 8
• Max memory file size: 100KB (~25K tokens)
• Version history retention: 30 days