Repository navigation
Secrets stay out of other agents' context.
Claude Code memories sometimes capture a credential from a debugging session. Served over MCP, that text would reach another agent's model provider. From 0.3.0, memory_get, memory_recall, memory_list and memory_due replace recognisable secrets with [redacted] before returning them, and the reply says how many were hidden. Files on disk are unchanged.
- Covered: passwords in connection URLs (scheme, user and host kept), bearer tokens, private key blocks, Anthropic, OpenAI, GitHub, Slack, AWS, Google and Stripe keys, and JWTs. Ordinary text (model ids, ticket numbers, password-less URLs) is left alone.
- Turn it off with
--allow-secretsorRECALL_ALLOW_SECRETS=1. - README: a tested Cline setup line next to Codex, Cursor and Gemini CLI (#2).
Suggested in openai/codex#58220. Changes: #1, #2.
Install or update:
npx -y @fainir/recall@0.3.0 --claude-code /path/to/your/repo --read-only
gemini extensions update recall