Remember the work. Protect the secrets.

Useful AI needs memory. It should know the project, the preferred format, the last decision, the team convention, the customer context, and the operating assumptions that would otherwise have to be repeated every session.

But memory is not the same thing as a Vault. Memory creates leverage. Vault protects trust. Treating them as one bucket is how useful context and sensitive secrets get mixed together.

A private AI system should remember how work happens. It should not casually remember API keys, SSH keys, access tokens, passwords, passcodes, private keys, customer secrets, or credentials. Those belong behind a stronger boundary with different interaction rules.

This distinction matters because the most helpful assistant is also the one closest to the work. The closer AI gets to files, tools, repos, systems, and decisions, the more precise the memory boundary has to become.

TARX’s model is simple: remember the work when the user approves it, keep secrets out of ordinary memory, and make the boundary visible enough that people can build trust without pretending all context has the same risk.