Source: Memory tool (Claude API docs)
Anthropic's reference for the client-side memory tool: a /memories file directory Claude can view, create, edit, rename and delete across sessions.
Contents
- What: official API documentation
- Publisher: Anthropic
- Date: read 2026-11-14 (tool version
memory_20250818, status stable) - Original: platform.claude.com/docs/en/agents-and-tools/tool-use/memory-tool
Summary
The memory tool gives Claude a directory of files under /memories that persists between conversations. It is client-side: Claude emits tool_use requests and the application executes them against storage it controls (disk, database, per-user prefix). The docs specify six commands, their reference return strings, the system-prompt instruction the API injects, security duties, and how memory pairs with context editing and compaction.
Key claims
- Configuration is a single tools entry,
{"type": "memory_20250818", "name": "memory"}; no beta header or input schema is needed. Available on Claude 4 and later models and on Bedrock, Google Cloud and Foundry. - Commands:
view(directory listing two levels deep, or file with line numbers and optionalview_range),create,str_replace(must match uniquely),insert(after line N, 0 = top),delete(recursive for directories),rename(must not overwrite). - The API auto-adds a system instruction telling Claude to view its memory directory first, record progress as it goes, and assume its context may be reset at any moment.
- Text views of files longer than 16,000 characters are truncated, so expect ranged follow-up views.
- SDK helpers:
BetaAbstractMemoryTool(Python, C#),betaMemoryTool(TypeScript),BetaMemoryToolHandler(Java);BetaLocalFilesystemMemoryToolships for Python and TypeScript. - Security is the application’s job: validate every path stays under
/memories(canonicalise, reject../and URL-encoded variants), cap file sizes, expire stale files, strip sensitive data before writing. - With context editing, Claude gets a warning before tool results are cleared so it can save what matters to memory first. With compaction, memory holds what must survive summarisation.
- Recommended multi-session pattern: an initializer session writes a progress log and feature checklist; later sessions read them first and update them before ending.
- Zero-data-retention and HIPAA eligible because storage is client-side; ZDR eligibility excludes Anthropic’s “Covered Models”.