Core concept: Claude Code reads CLAUDE.md from a small, fixed set of places on the local filesystem. Two of them are your home directory and directories inside the project. It never loads instructions from a remote object store.

Why A and C are correct:

A: The memory docs list "User instructions" at ~/.claude/CLAUDE.md, whose purpose is "Personal preferences for all projects." It applies in every repo you open.
C: Subdirectory files are supported too: "Claude also discovers CLAUDE.md and CLAUDE.local.md files in subdirectories under your current working directory." They "are included when Claude reads files in those subdirectories," so src/api/CLAUDE.md scopes conventions to that module and costs nothing until it is needed.
Both are layers a developer creates by hand. The project root file, ./CLAUDE.md, is the third everyday one.

Why the others are wrong:

B: A machine-wide file does exist, but not there. The managed-policy path is /etc/claude-code/CLAUDE.md on Linux and /Library/Application Support/ClaudeCode/CLAUDE.md on macOS, deployed by IT. /etc/claude/CLAUDE.md is read by nothing.
D: There is no cloud layer and no s3:// support. Memory files load from disk; you share them through source control.

Remember: CLAUDE.md is local and hierarchical: home directory, project root, then any subdirectory.

Question 160