Zero Data Retention Architecture
A technical breakdown of OpenRouter account-level ZDR enforcement, provider-level zero retention agreements, and ephemeral inference.
The Core Privacy Guarantee
The defining security principle of ARMES is Zero Data Retention (ZDR) at the inference layer.
Traditional consumer AI interfaces store every prompt, attachment, and response in provider databases for 30+ days or longer. Many providers reserve the right to review transcripts with human contractors or use your conversations to train future foundation models and monetize user data.
In ARMES, every model request is routed through OpenRouter with Zero Data Retention enforced at the account level. Transport is encrypted (TLS); the retention guarantee itself is contractual, not cryptographic.
Account-Level ZDR Enforcement
ARMES routes all model compute strictly through enterprise infrastructure managed by OpenRouter with zero-data-retention and data-collection restrictions enforced both at the account level and on every runtime request (zdr: true, data_collection: "deny").
Crucially, we do not use consumer AI lab endpoints. When you converse with frontier models in ARMES, requests are routed exclusively to enterprise hyperscalers:
Enterprise Endpoint Routing:
ARMES never picks a provider endpoint directly. OpenRouter selects one from the endpoints that meet our account-level Zero Data Retention requirements (such as Microsoft Azure for OpenAI and Amazon Bedrock for Anthropic). Your prompts are processed ephemerally to produce the response and are not stored, logged, or used for training.
Non-ZDR Model Disqualification Policy
ARMES strictly excludes any model that fails to offer enterprise Zero Data Retention agreements. For example, consumer or experimental releases that lack upstream enterprise zero-retention compliance or mandate data collection for internal testing are strictly disqualified from the ARMES model catalog, API routes, and auto-router pools. If an upstream model cannot guarantee processing without disk persistence or human oversight, it will never be added to ARMES.
Why ZDR Goes Far Deeper Than "No Training"
Many AI products claim "we do not train on your data," but still retain conversations on disk for 30 days or longer to allow human contractor review or law enforcement lookups.
ARMES enforces Zero Data Retention (ZDR) at rest, which answers three distinct architectural questions:
- Zero Provider Retention: Under ZDR terms, the enterprise provider processes incoming tokens to generate the completion and does not store prompts, documents, or completions on persistent storage afterward.
- Zero Model Training: Because no data is retained, your conversations, memories, files, and project contexts cannot be used to train, fine-tune, or benchmark any future foundation model or reward checkpoint.
- Zero Human Review Queues: Under OpenRouter ZDR enforcement, any endpoint that retains logs for abuse inspection or human contractor moderation is classified as non-ZDR and strictly excluded from our routing pool.
- Zero Ad Profiling & Data Brokering: Your queries and user metadata are never linked to advertising identifiers, cross-site trackers, or commercial profiling systems. Revenue is 100% funded by subscriptions and transparent credit top-ups.
Dual-Layer Privacy Invariant
To understand your data boundaries, remember the Brain vs. Vault distinction:
External Cloud Storage Boundaries
When you connect third-party cloud services like Google Drive:
- Google Terms Apply: Documents in your Google Drive reside on Google servers governed by Google's privacy terms. ARMES does not store these files in the Vault unless you explicitly choose to import them.
- Encrypted Token Broker: ARMES encrypts your OAuth credentials with AES-256-GCM and never exposes raw credentials to AI models. Agents interact with external files through an ownership-scoped broker.
- Prompt Injection Quarantine: Text read from external documents is automatically wrapped in
<external_data>quarantine tags and capped at ~30k tokens to protect LLM context windows against prompt-injection attacks. - Action Proposal Gateway: Agents can draft modifications, but dangerous overwrites are intercepted and require your explicit interactive confirmation before any external file is altered.
- Full Auditability: Review all agent reads, searches, and edits in the Integration Activity & Audit Log in Settings.
Verification & Auditing
You can audit data handling at any time:
- Inspect our live terms at armes.ai/privacy and armes.ai/security.
- Model routing logs show transparent timestamps and token tallies without persisting message content.