01 The trust boundary
A diagram and a promise: code, prompts, and decisions stay inside your perimeter unless you explicitly cross it.
02 Encryption
Industry-standard, no clever inventions.
03 Isolation
A workspace is a hard boundary. No accidental cross-talk.
- Per-workspace encryption keys. A workspace can be re-keyed independently. A leaked key for one workspace doesn't compromise another.
- Per-workspace embedding namespaces. Even if vectors collided across workspaces, the namespace filter would prevent retrieval.
- Per-workspace Postgres schema (Enterprise). No shared tables, no shared connection pool.
- Network policy. On VPC/Enterprise, namespace-level NetworkPolicies prevent cross-pod traffic.
04 Access control
RBAC, scoped tokens, SSO, SCIM. Authentication is boring; authorization should be sharp.
owner · admin · maintainer · contributor · read-only. Custom roles on Enterprise.05 Audit log
Every read, every write, every grant. To your SIEM, in real time.
- Immutable, append-only chain. Each entry signed; tampering is detectable.
- Streams to Splunk, Datadog, Sumo, syslog, or generic webhook (Enterprise).
- Retention: 90 days in-app on Pro; up to 7 years configurable on Enterprise.
- Filterable by actor, workspace, scope, time, action, result. Exportable as CSV/JSON.
- Includes citation chain: when an agent uses a memory, the audit links the prompt → retrieval → cited source.
06 DLP & redaction
Strip the things that shouldn't be remembered before they're stored — not after retrieval.
07 Right to forget
Delete means delete.
- Hard delete propagates across primary store, vector index, graph, hot cache, and cold archive within 24 hours.
- The deleted item is gone; an audit entry "an item with this hash was deleted" remains for compliance.
- Workspace deletion takes the workspace's encryption key with it. Cryptographic shredding closes the loop on cold backups.
- Self-hosted operators can run
statefulai forgetfor explicit, scriptable removal.
08 Models & data
Your code is not training data.
- We do not use your data to train or fine-tune any shared model. Ever.
- Per-workspace rerankers are trained on your accept/reject signals; their weights stay in your workspace.
- On Enterprise, you can bring your own embedding provider — Voyage, Cohere, OpenAI, or a self-hosted model.
- Embedding requests to third parties are content-redacted before they go out (when you opt in to a hosted embedder).
09 Vulnerability disclosure
Found something? Tell us, get paid, and we'll thank you publicly.
security@statefulai.tech — PGP key on file.