Infinity Neural Memory exists partly because a predecessor leaked plaintext secrets. We take security seriously and bake it into the code (inmem/redaction.py, the AdmissionGate), but defense-in-depth still needs you.

Do not open a public issue for security problems.

Please report privately via GitHub Security Advisories ("Report a vulnerability" on the repository's Security tab), or by email to the maintainer listed in pyproject.toml.

Include: affected version, reproduction steps, and impact. We aim to acknowledge within 72 hours and to ship a fix or mitigation for confirmed issues as a patch release.

• The timeline (timeline/) and store (*.db) are private by default. They may contain user content. Never commit them — .gitignore and the CI gate both block this.
• Redactor scrubs common secret formats (API keys, tokens, JWTs, private keys, password key/values) on every write path and keeps only a SHA-256 of the original. This reduces risk; it is not a guarantee. Do not feed known secrets into memory intentionally.
• The AdmissionGate blocks writing any content in which a secret is detected (block_sensitive=True by default). Keep it on.
• Run the secret scanner in CI (gitleaks) and locally (pre-commit). Both are pre-wired.

In scope: accidental secret persistence, accidental commit of private memory, basic duplicate/poisoning hygiene via the AdmissionGate.

Not yet in scope (roadmap): adversarial memory-poisoning defense, per-source trust scoring, low-confidence recall rejection, and cryptographic erasure / right-to-forget. Track these in references/DESIGN.md → Roadmap.