Empirica measures what your AI agents know, so it sits close to sensitive work. That
means you decide where your knowledge lives.
Run a practice with no Empirica account and nothing leaves your machine. Connect one
and your work syncs to your own org workspace for backup and shared memory, never across
an organization boundary without your explicit agreement, and anything you tag
local never leaves your device at all. These aren't
policies we promise; they're structural properties you can verify in the source.
🏠 Where your data lives
A practice, meaning its knowledge graph and its artifacts (findings, decisions, unknowns, dead-ends), always writes to a local SQLite database and a local vector store on your machine first. Whether a copy leaves comes down to one choice:
- No Empirica account connected: nothing egresses. Your artifacts live only in those local stores. Fully on-device.
- An Empirica account connected: by
default your artifacts sync to your own org workspace,
so you get backup, memory across sessions, and discovery by your team's other
practices. Tag an artifact
visibility=localand it is never transmitted to Empirica servers, even with an account connected.localmeans on-device, full stop. No other organization can read your data without an explicit, mutual sharing agreement.
Either way the content is yours; the only question is whether a copy also lives in your workspace. A live mesh session additionally sends lightweight coordination signals: presence (which practice is active, on which machine, its status) and per-session calibration vectors (13 numeric values, never the text of a finding or decision).
🔀 What the coordination layer does
Empirica's cortex layer coordinates the mesh: it routes proposals and collaboration messages between practices, tracks which practices are present, and emits system events like delivery failures. It also holds your own workspace when you connect an account, as above.
The visibility tag controls sharing. An
artifact tagged local is yours alone: never shown
to another practice, never promoted to org-wide or public discovery. Marking one
shared (your org) or public
(mesh-wide) is an explicit, per-artifact decision to let others find it. And
local does double duty: an artifact tagged
local is never shared with another practice or org,
and it is never transmitted to Empirica servers at all, even with an account connected.
💾 Backup is opt-in, and it's just git
Nothing is backed up automatically. Backup is opt-in and git-based (Empirica syncs artifacts and session history through git-notes), and you point your practice at any git remote you control. If you'd rather not run your own, Empirica can host a private git repository per practice on request, but there is no default backup destination, and no data leaves until you elect one.
🖥️ Where practices run
Practices are meant to run on your own infrastructure: your laptop, your servers, your cloud. Empirica does operate a hosted execution surface, but it's available only by explicit arrangement; it is not a self-serve product today. In other words: running your practice in someone else's cloud is a thing you'd have to ask for, not something that happens by default.
🔐 Security posture
Encryption
HTTPS/TLS for all API and coordination endpoints. Coordination-layer secrets (tokens, per-project keys) are encrypted at rest with a key held out-of-band from the database. On your machine, your disk encryption (FileVault / LUKS / BitLocker) is the control for local stores.
Tenant isolation
Every artifact and proposal is scoped to an (org, user, project) identity at write time. Cross-tenant reads require an explicit, revocable sharing agreement that both sides opt into, enforced in production. One org's key cannot read another org's data.
Retention
Coordination messages carry a 30-day server-side lifetime, then archive. Your
artifacts are durable until you delete them (empirica delete-artifacts).
Presence and session records are durable and can be deleted on request.
🛡️ Built for the regulations, and continuously proven
Empirica is engineered against the EU AI Act, ISO/IEC 42001, and GDPR control frameworks from the ground up. Every release passes an automated compliance pipeline that maps 13 quality, security, traceability, and human-oversight controls directly to their regulatory obligations: record-keeping (Art. 12), human oversight (Art. 14), transparency (Art. 13 and 50), accuracy and robustness (Art. 15), and data governance (Art. 10).
Our epistemic audit trail, human-in-the-loop decision gating, and calibration measurement are not retrofitted for compliance; they are the architecture. The pipeline runs on every push (linting, type-checking, a prompt and parser drift check, the test matrix, a weekly dependency and CVE scan, and our quality-audit skills), and today all 13 controls pass. This is verifiable engineering rigor, continuously enforced in CI, not a badge.
📋 Compliance: the honest version
Empirica does not hold SOC 2, ISO 27001, HIPAA, or a formal GDPR certification. We're a small team, and formal audits aren't on the near-term roadmap. We'd rather tell you that plainly than imply otherwise.
What we do offer is structural privacy by design: local-first storage and tenant isolation are properties of how the system is built, verifiable in the source rather than asserted in a badge. If you're a regulated buyer with a specific requirement, talk to us and we'll be straight about whether Empirica fits.
Don't take our word for it: the local-first design is in the open source.