How Claude remembers your project
Source: Claude Code Docs Topic: Persistent instructions and auto memory in Claude Code.
1. Two memory layers
Claude Code distinguishes between user-authored CLAUDE.md files and automatically collected memory. CLAUDE.md carries explicit project, user, or organization instructions; auto memory stores learnings and patterns Claude discovers while working.
2. Scoped instruction files
The docs define multiple CLAUDE.md scopes, including managed policy, project instructions, and user instructions. More specific files override broader ones, which gives Claude Code a structured hierarchy for persistent guidance.
3. Cross-session carryover without pretending the transcript is infinite
Anthropic’s framing is pragmatic: Claude reads these files at session start, and the docs link this memory model to session management, context exploration, and resume flows. The persistence model is therefore artifact-based rather than raw transcript accumulation.
4. Subagent memory
The page also notes that subagents can maintain their own auto memory. That makes memory isolation part of orchestration, not just a global shared store.