Integration pattern: applying the reference architecture in VeraMem
VeraMem is the author's own host application for ARVIS, not an independent adoption: read this as an illustrative integration pattern, not as a case study with deployment evidence (it deliberately contains no dates, scale, metrics or incident history). It explains the architectural relationship at a high level so an integrator can project ARVIS into a real system. It is descriptive, not a public contract for VeraMem, and does not expose production configuration, credentials or proprietary business logic.
Read the generic governed-assistant reference architecture first.
Responsibility split
| VeraMem host responsibility | ARVIS responsibility |
|---|---|
| user identity, sessions and personal/organization spaces | validate the trusted identity material presented to an effect |
| document ingestion, OCR, metadata and access rules | govern whether a proposed operation may proceed |
| relational and vector persistence | emit canonical IR and commitments |
| retrieval and prompt/context assembly | compose policy, uncertainty and admissibility signals |
| model-provider selection and credentials | treat model output as untrusted proposal data |
| queues, workers, retries and crash recovery | bind an authorized payload to a single-use capability |
| consent UI and recipient resolution | require and bind confirmation where policy demands it |
| business tools and external connectors | enforce the governed syscall and audit sequence |
VeraMem remains the product and integration layer. ARVIS remains the application-independent governance kernel.
For a context-dependent turn, VeraMem may set
host_context_available=True only after it has loaded, authorized, selected
and judged relevant the product-side context. The flag is a content-free
attestation: no transcript, document passage or prompt fragment enters ARVIS
through it, and it does not become long-term memory. ARVIS uses the signal only
to distinguish "context absent" from "relevant host context is available" when
producing its contextual uncertainty gap.
Illustrative topology
flowchart TD
UI["Web or self-hosted UI"] --> API["VeraMem API"]
API --> GOV["ARVIS"]
API --> MEM["SQL and vector memory"]
API --> AI["Local or declared model"]
GOV --> WORK["Governed workers and tools"]
Concrete technologies can include PostgreSQL, Qdrant, RabbitMQ and a local model runtime such as Ollama. They are VeraMem deployment choices, not ARVIS requirements.
Example flow
Consider a user asking:
Find the latest contract note, summarize the termination clause and send it to the counsel working on this matter.
VeraMem:
- authenticates the user and selects the active matter or space;
- applies document access rules before retrieval;
- retrieves source passages and constructs model context;
- resolves the intended recipient through trusted application data;
- translates the proposed send into a typed tool request;
- displays any ARVIS-required confirmation with the exact recipient and bounded content;
- supplies the durable audit and idempotency infrastructure used by the governed effect.
ARVIS:
- governs the structured decision;
- distinguishes the read-only retrieval from outbound data egress;
- refuses malformed or unsupported inputs fail-closed;
- binds the authenticated principal, tenant/space and payload;
- prevents execution until policy and confirmation conditions hold;
- returns the decision, trace, IR and commitment needed for audit or replay.
The LLM never receives authority to send the message. The connector never receives mutable cognition state. The product host never receives a sanctioned shortcut around the effect path.
Why publish this pattern
ARVIS is intentionally abstract. VeraMem demonstrates that its boundaries map onto familiar product concerns:
- a memory system remains responsible for retrieval and storage;
- a model remains responsible for generation;
- an application remains responsible for users, spaces and business rules;
- ARVIS governs the transition from proposal to allowed effect.
This makes the kernel easier to adopt without coupling it to one database, model provider or product.
Relationship to a future self-hosted edition
A planned self-hosted VeraMem edition can package the application pieces that ARVIS deliberately does not provide: UI, accounts, spaces, memory, deployment configuration and ready-made connectors. ARVIS remains independently useful for developers who want to build a different host.
This creates two distinct adoption paths:
- use ARVIS directly as a Python governance kernel;
- use VeraMem as a more integrated assistant product built around it.
Neither path changes the security boundary: production identity, durable audit, consent, provider configuration and operational recovery remain explicit host obligations.
Deliberate omissions
This document does not publish:
- VeraMem production endpoints or network topology;
- secrets, certificate policy or deployment credentials;
- database schemas or tenant-isolation implementation details;
- proprietary ranking, prompt or product logic;
- a production-ready compose file.
Those details are unnecessary to understand where ARVIS fits and would turn a reference architecture into an accidental second VeraMem distribution.