What Is Claude Memory?

The short version

Claude can remember across sessions, but the mechanism depends on where you are using Claude. In the Claude app, memory is a set of short topic files built as you chat, and since September 2026 the same memory is used by Claude Cowork. In Projects, memory is separated by project. In Claude Code, memory usually means persistent instruction files and auto-written Markdown notes. On the developer platform, memory can mean an agent tool that stores files outside the context window.

The practical mistake is asking "does Claude remember?" without asking "which Claude surface, which scope, and who controls the storage?" A useful memory setup begins by naming the surface first. For the current chat and Cowork behavior, see memory everywhere.

The five memory surfaces

Chat and Cowork memory is the user-facing topic list Claude builds as you chat, shared between conversations and Claude Cowork tasks. Use it for personal work context, stable preferences, and recurring collaboration details.

Project memory keeps a project's summary separate from non-project chats. Use it when the memory should not leak into unrelated work.

CLAUDE.md is written by a human and loaded as Claude Code context. Use it for rules, conventions, architecture notes, and commands that should be present every session.

Auto memory is written by Claude Code itself as plain Markdown. Use it for learned preferences and project facts Claude discovers while working.

Platform memory tools are application-level storage patterns. Developers decide where the data lives and how Claude reads or writes it.

Memory is context, not enforcement

Claude Code documentation is explicit that memory files are context rather than hard configuration. That matters. A memory can influence behavior, but if a workflow must run at a fixed time, such as before a commit or before a tool call, use a hook, policy setting, or automation rather than a reminder in a memory file.

This distinction explains many "Claude ignored my memory" reports. Persistent context improves odds; it does not make a soft instruction into a guardrail.

What to read next

If you use chat, Cowork, or the mobile app, start with memory everywhere for what changed in September 2026, then Claude app memory for the day-to-day controls. If you use the terminal, read Claude Code memory and then the deeper guides to CLAUDE.md and auto memory.

For context-window sizing, compaction, and prompt caching, use the sibling resource Claude Context. For broader Claude Code engineering practice, use Claude Ships Code.