I didn't realize Claude Code was logging my entire machine until the leak

A March 2026 accidental source code leak from Anthropic revealed exactly what Claude Code collects from users' machines: not just prompts, but the actual contents of any files the tool reads, plus memory data stored on Anthropic's servers. Data retention can range from 30 days to seven years depending on plan tier and safety flags. The leak also exposed unreleased background features like 'dream mode' (idle reflection on past interactions) and KAIROS. None of this technically contradicts Anthropic's existing terms, but the practices were scattered across policy pages most users never read. Anthropic has since published environment variables to disable memory and telemetry, including CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 and a --bare mode flag.