Mnemosyne is a personal/lab project. If you find a security issue, please report it privately via GitHub Security Advisories rather than opening a public issue. You'll get an acknowledgement and a fix timeline.

Mnemosyne is local-first by design: it runs against a local Ollama instance and a file-based FAISS index, with no cloud LLM calls and no API keys. That removes a large class of risk, but a few things are worth keeping in mind:

• FAISS deserialization. Loading an index uses pickle (allow_dangerous_deserialization=True). Only load indices you built yourself. Treat a FAISS index from an untrusted source as untrusted code.
• Ingested documents. Anything you ingest becomes context the model may surface verbatim. Don't ingest secrets you wouldn't want echoed into an answer.
• URL fetching. ingest can fetch URLs listed in a pack's sources.yaml. Only list URLs you trust; fetched content is treated as a document, not executed.
• Third-party docs / licensing. Mnemosyne ships no vendor documentation. Keep ingested corpora out of version control (the knowledge/ path is gitignored).

Dependencies are monitored via Dependabot and patched through normal PRs.