Loading...

Security & Sovereignty

Your data. Your environment. Your control.

Mnemo is built on a simple security principle: your documents and queries should never leave your control. This page covers the specific technical and architectural measures we use to enforce that guarantee. Written for IT security leads and Data Protection Officers.

Core security guarantees

These are not aspirations or roadmap items. They are architectural properties of Mnemo as deployed today.

Zero data exfiltration

Document content, query text, and AI responses are processed entirely within your deployment environment. No document data is transmitted to external AI providers. The AI inference layer operates within the same security boundary as your data.

EU/UK sovereign deployment

Mnemo is deployed in EU-based infrastructure. UK-only deployment is available. We do not use multi-region replication that would move data across jurisdictional boundaries without your explicit agreement. Jurisdiction is agreed in writing at contracting.

Customer-controlled encryption keys

You can supply your own encryption keys for data at rest. This means Arion Flow cannot access your data without your keys — even in an administrative context. Key management is handled via your preferred KMS provider.

Multi-tenant row-level isolation

All customer data is logically isolated at the row level using enforced tenant identifiers. Database queries are scoped to the authenticated organisation at the query engine level — not at the application layer. Cross-tenant data access is architecturally prevented.

SHA-256 blob hashing

Every document stored in Mnemo is hashed using SHA-256 on ingestion. This hash is stored and verified on every retrieval. Any file modification — including partial tampering — is detectable. This supports data integrity verification for audit and legal purposes.

No training on customer data

Your documents and queries are never used to train, fine-tune, or update any AI model. This applies to the models used within Mnemo and to any third-party model providers in the processing chain. It is a hard contractual and architectural commitment.

How the security architecture works

Mnemo uses a retrieval-augmented generation (RAG) architecture. Understanding how data flows through this architecture is essential to understanding the security model.

1

Document ingestion

Documents are uploaded or connected via secure connectors (SharePoint, S3). They are processed — chunked, vectorised, and hashed — entirely within your deployment environment. The original files are stored in your controlled storage layer.

2

Vector indexing

Document chunks are converted to vector embeddings using models running inside your environment. These embeddings are stored in a vector database scoped to your organisation. No embedding data leaves your environment.

3

Query processing

When a user submits a query, it is vectorised and matched against your index. The matching chunks are assembled into a context payload and passed to the language model. All of this happens within your security boundary.

4

Response generation

The language model generates a response based only on the retrieved context — your documents. It does not draw on general internet knowledge or cross-customer data. The response is returned with source citations.

Access and authentication

Authentication

OAuth 2.0 via your identity provider (Google Workspace, Microsoft Entra ID) or email/password with TOTP available.

Authorisation

Role-based access control (RBAC) at workspace level. Roles: Organisation Admin, Workspace Member. Permissions enforced at the API layer.

Session management

Sessions expire after a configurable idle period. All sessions are invalidated on password change or account suspension.

API access

API keys are scoped to a workspace. Keys can be rotated or revoked independently. All API calls are logged.

Data in transit

All data in transit is encrypted using TLS 1.2 or higher. Internal service-to-service communication within the deployment environment is also encrypted. There are no unencrypted data paths.

Data at rest

All stored data — documents, vector embeddings, user data, audit logs — is encrypted at rest using AES-256. Customer-managed keys (BYOK) are available for organisations requiring key custody.

Information for Data Protection Officers

Key facts relevant to GDPR Article 30 record-keeping and data protection impact assessments.

Data controller

Your organisation

Data processor

Arion Flow Ltd (UK registered)

Processing location

EU by default; UK-only available

Legal basis for processing

GDPR Article 6(1)(b) — contract performance; Article 6(1)(f) — legitimate interests

Retention period

As agreed in the DPA; default is duration of service + 30 days

Sub-processors

Listed in the DPA; 30 days' notice of changes

Data subject rights

Supported via data export and deletion tools; response within 72 hours of request

Breach notification

Notification to your organisation within 24 hours of confirmed breach; ICO notification support available

Want a detailed security briefing?

We can arrange a technical conversation with our engineering team for organisations requiring deeper assurance — penetration test results, architecture diagrams, and sub-processor details available under NDA.

Request a Security Briefing