Memory shipping

Just files.

Memory is a folder of human-readable Markdown notes under ~/.claude/projects/<dir>/memory/. An index file (MEMORY.md) is loaded into every conversation; individual memory files are pulled in when relevant. You can read, edit, and version-control everything by hand.

Save the right things.

Memory ≠ documentation.

• Code patterns, conventions, file paths — derive from the project state.
• Git history, recent changes — git log is authoritative.
• Debug fixes — the fix is in the code; the commit message has context.
• Anything in CLAUDE.md.
• Ephemeral state — current task, in-progress notes.

Keep it from rotting.

• Verify before recommending. A memory says "old_function exists" — grep before recommending it. Files get renamed.
• Update when wrong. If a recalled memory conflicts with reality, fix the memory, don't act on it.
• Index discipline. MEMORY.md entries are one line each, under 200 chars. Detail goes in topic files.
• De-dupe. Before writing a new memory, check if an existing one can be updated.

Bring it to the SDK.

The Agent SDK exposes the same memory primitives. Hand it a memory directory; it reads on start, writes throughout the run. Useful for long-running agents that revisit the same projects across sessions.

Related products.

Every Anthropic surface has its own page on this guide.