Mnemolith

Team memory never shares what it shouldn’t.

How team memory keeps keys, passwords and your own memories out of your repo. Nothing here involves a server: there isn’t one. Last updated November 13, 2026.

What goes into your repo

Only the memories someone on your team chose to share, each as a small text file in .mnemolith/team/memories/, plus a one-line list of them in mnemolith.md and one line in .gitattributes. Each file holds the fact, who wrote it, when it was confirmed, and the names and fingerprints of the files it depends on. Never the contents of those files.

Mnemolith writes these files and stops there. It never commits or pushes: they leave your computer only when you push, to wherever your repo already lives, and your team reviews them like any other change.

What never goes into it

Your own memories, unless you share one. Mnemolith’s data, settings and model in ~/.mnemolith, including which repos you use team memory in and your sharing choices. Your code, your prompts and your conversations with your AI.

The secret check

Every memory is checked on your computer before it’s written to the repo: when it’s shared, reworded, imported from your notes, or drafted by your AI. The check refuses anything that looks like a private key, a cloud or API key, an access token (GitHub, GitLab, Slack, Stripe, npm, PyPI, Hugging Face and others), a password written out (DB_PASSWORD=…, or “the admin password is …”), a login and password together, a web address with a password in it, or a card number. When it refuses, it shows only the first few characters, so the secret isn’t printed in full.

Something that only might be a secret, like a long random-looking string or a folder path with your user name in it, is left for a person to decide. Your AI can’t wave it through. And mnemolith.md leaves out any memory that fails the check, even one typed into a file by hand.

The check is a set of patterns, not a guarantee: it catches the common shapes, and review in the pull request is the second line of defence. If a secret ever does get committed, change the secret. Deleting the file doesn’t remove it from git’s history.

Who can see team memories

Everyone who can see the repo, and no one else: team memory uses your repo’s own permissions, with no separate accounts. If the repo is public, that’s anyone on the internet, which is why Mnemolith asks before every share, unless you turn that question off.

What AI tools can do

Read team memories in recall once you’ve joined with mnemolith team init. Cloning a repo isn’t enough. Share a memory only after you’ve allowed it with mnemolith team ai on, and only when you ask. They can never commit, and never get past the secret check.

If you ask Mnemolith to draft starter memories, it runs your AI tool once on your computer, under your own account with that tool, kept to reading files and adding team memories as tightly as each tool allows: Claude Code gets an explicit list of allowed tools (no edits, no commands beyond git log and ls), Codex runs in its read-only sandbox, and Cursor with its sandbox on. For Cursor, Mnemolith points its own settings file at Mnemolith for that run and puts the file back exactly as it was. If a tool changes any other file anyway, Mnemolith lists them so you can undo them.

Found a problem?

If you find a way to get a secret past the check, or anything else that puts data at risk, please report it privately to [email protected], or through a private security advisory on GitHub, rather than in a public issue.