This project is an educational prototype. The points below describe a practical hardening baseline for handling sensitive health-adjacent data.

• Data minimization: only collect fields required for risk scoring and explainability.
• No raw identifiers by default: avoid name, email, phone, address, national IDs in request payloads.
• Redaction in logs: audit logs should include only normalized clinical features and decision metadata.
• Need-to-know access: restrict memory/history endpoints behind token auth in non-local environments.
• Environment separation: keep development/test data isolated from demo/production-like data.

• Default retention (prototype): keep decision/audit history for 30 days.
• Memory retention: rotate or archive old ChromaDB patient vectors on a schedule.
• Right to delete: support deleting patient-associated memory records on request.
• Backups: encrypt backups and define retention windows for snapshots.

• Enforce API auth (REQUIRE_API_TOKEN=true) and rotate tokens regularly.
• Move from shared token auth to per-user auth (JWT/OAuth) for multi-user environments.
• Use HTTPS termination and secure secrets storage (vault/KMS, not plaintext env files).
• Add rate limiting and request size limits at API gateway/reverse proxy level.
• Add structured security events and alerting for auth failures and unusual access patterns.

• Encrypt data at rest for vector store, logs, and backups.
• Encrypt data in transit between UI, API, and storage.
• Use separate credentials/roles for read-only analytics vs write paths.
• Prevent accidental commits of logs/databases by keeping them in .gitignore.

• Treat this as non-diagnostic decision support.
• For regulated use, complete threat modeling, DPIA/HIPAA review, and clinical validation.
• Maintain auditability of model version, feature inputs, and recommendation provenance.