When Claude Code is in interactive mode you can have Claude modify its own CLAUDE.md temporarily to run tasks with a specific context. After the task is completed you can ask Claude to revert the CLAUDE.md back to its previous state. There are several ways of backing up the CLAUDE.md:

• Git versioning: Use git stash or commit before changes for easy rollback.
• File duplication: Have Claude copy to a backup filename like CLAUDE.md.backup.

To change the information persisted through the CLAUDE.md interface you can use 'quick Memory' with #.

This mechanic is useful if you have a CLAUDE.md library which specifies the processes for performing different tasks. Claude is unable to read changes to the CLAUDE.md into memory mid-session unless it utilises explicit commands such as:

• Quick Memory: Using # commands to temporarily change persisted information.
• Memory refresh commands: Explicitly asking Claude to re-read the modified CLAUDE.md.
• Session restart: Starting a new Claude Code session to pick up changes.
• Explicit file reads: Using read commands to load the updated content.

This led to me exploring getting Claude to spawn another instance of itself in a directory which has a different CLAUDE.md. The benefits being:

• Cleaner context separation: Each spawned instance operates with its own specific CLAUDE.md context.
• No memory reload issues: New instances automatically load their directory's CLAUDE.md.
• Parallel processing: Multiple contexts can run simultaneously without interference.
• Reduced context pollution: Previous session context doesn't bleed into the new specialized task.

Note: Claude instances will automatically crawl up the directory structure and read any CLAUDE.md files they encounter. So to have a strict separation of context ensure the project directory CLAUDE.md is lightweight and non-specific.

See Also: CLAUDE.md Supremacy|Context Window Constraints|Context Window Depletion