nmemory's security posture is structural: the binary opens no network socket (compiled without a networking stack), touches one SQLite file on your disk, and treats everything it returns as data, never instructions (ADVISORY_NOT_AUTHORITY framing on every envelope). The threat that matters most here is memory poisoning — stored content trying to become a command — and the armor is the framing contract, not a detector.

Use GitHub private vulnerability reporting: the Security tab of this repository → Report a vulnerability. That opens a private advisory only the maintainer can see.

Please include: the version (nmemory --version), a minimal reproduction, and what an attacker gains. A capture/recall payload that escapes the ADVISORY_NOT_AUTHORITY framing, bypasses provenance enforcement, or makes recall fabricate an outcome is in scope and taken seriously — that class is the project's reason to exist.

Do not open public issues for exploitable findings before a fix ships.

• Attacks requiring write access to the store file itself (your disk, your trust boundary — protect the file like any local secret).
• Prompt-injection resilience of the consuming agent: nmemory guarantees its own output framing; what an agent does with framed data is that agent's boundary.