Persistent memory lets a subagent carry state forward between invocations. A researcher agent can remember what it's already looked into; a reviewer can remember your house style.

What it does

Enable persistent memory in the subagent frontmatter and Claude writes summaries to a scoped memory file after each run. On the next invocation, the agent loads those notes before it begins. It's auto-memory, but scoped to the agent identity rather than the project.

When to use it

Agents that do the same job repeatedly and benefit from recall.
Multi-session work where you want continuity without copy-pasting context.
Reviewer or auditor agents that should learn your preferences over time.
Fan-out patterns where each worker builds on prior runs.

Gotchas

Memory grows over time. Prune stale entries or it starts to slow down invocations.
Persistent memory isn't magic - the agent still needs relevant instructions each call.
Scope matters. Project-scoped memory shouldn't leak personal preferences; user-scoped shouldn't leak project data.

Official docs: https://code.claude.com/docs/en/sub-agents.md#enable-persistent-memory