Trust & sovereignty
Your data has
one custodian. You.
Every sovereignty claim on one page — the data boundary, provenance, audit, models, and what happens if we disappear. Each claim labeled shipped, engine-ready, or designed.
The data boundary
What never leaves
your machine.
- Your Brain — the Spectral database — lives on your disk
- Your documents and files stay where they are
- Your memories, skills, and preferences never upload
- Your API keys — never relayed, never stored, never seen by us
Model calls go directly from your machine to your chosen provider, on your account — no Permagent server in the middle, no second copy of your data. Beyond that, no analytics beacon unless you turn it on — off by default — and the in-app egress audit shows you exactly what left.
And what does — only when you ask
- Inference
- Your prompt goes to the provider you chose, on your account, directly — no Permagent server in the middle and no second copy. Choose a local model and even this stops.
- Embeddings
- Only if you enable a provider that needs them. Spectral recall itself is deterministic fingerprinting, so the default path computes no embeddings at all.
- Telemetry
- Off on a fresh install. No beacon unless you turn it on, and it is slated to move off third-party SaaS.
All three are gated by the engine’s egress guard and recorded in the in-app audit view. In sovereign mode the guard is fail-closed: the call fails before any data leaves the machine.
The guarantees
Sovereignty, itemized.
Each card is a claim we stand behind, and each carries one of three labels. No vaporware sold as features.
Provenance
The Spectral engine records which brain asserted each memory, under a per-brain Ed25519 identity. The app doesn't stamp author identity on writes yet — that lands with team sync.
Scoped visibility
Four visibility scopes — Private, Team, Org, Public — exist in the engine. Today the app writes everything Private and there is no in-app scope control yet; that arrives with team sync.
Model-agnostic
Anthropic, OpenAI, Bedrock, Vertex, and fully local inference through Ollama and llama.cpp. Direct connection, your account, no middleman markup.
Open formats
Skills and memory export to plain files in one command. If you ever want to leave, you walk out with everything.
Open source
The code is open. The data format is yours. If we disappeared tomorrow, your agent — and everything it knows — keeps working.
No beacon by default
A fresh install sends zero analytics — no beacon unless you turn it on. Opt-in telemetry exists if you want it, and is slated to move off third-party SaaS.
Skill run history
Skill runs are recorded locally — auto-detected skills and directly-run saved skills alike — so you can see what a skill did and when.
Egress & audit log
An in-app egress audit view shows exactly what left the machine — inference, embeddings, telemetry — gated and recorded by the engine's egress guard.
Forget, verified
A verified hard-delete primitive in the engine — per-key and per-wing. Not yet exposed as an in-app control. Scrubbing graph-derived facts is planned.
For the team buyer
The trust model scales
to groups.
Provenance and scopes are the audit trail a shared team Brain runs on: every fact carries its contributor, every memory carries who may see it. Trust is a property of the memory model, not a console setting.
Team federation — shared Brains with member-controlled scopes and clean, auditable offboarding — is the next build on the engine already in v1.31.
What we don’t claim
No compliance certifications yet
We hold no SOC 2 or ISO 27001 certification today, and we won’t imply otherwise. If that changes, it changes in the build log first.
Team sync is not live
Four visibility scopes, identity, and provenance ship in the engine — but the app writes everything Private, has no scope controls, doesn’t stamp author identity, and doesn’t sync anything between members yet. Team sync is the next build.
Local-first, not magic
Sovereignty means custody and auditability — it doesn’t make the model calls you choose to send private from your provider, and it doesn’t stop a person from copying what they can already see. We’ll always be precise about that boundary.
Sovereignty you can verify.
Open code, open formats, your disk, your keys. Join the waitlist to get the first public build.