Canonadr

ADR-2026-09-04-14 — Four-Flow Durable Business Context, Multi-Modal Evidence, and Evolutionary Memory

Status: ACCEPTED — 2026-09-04 (founder core thesis; four-flow ingestion; memory & evidence preservation)

Context

A recurring failure mode in AI integration architectures is reducing the platform to an ephemeral tool-proxy, metadata catalog, or pure vector-RAG store. When connected to host environments (e.g. ChatGPT, Claude, Slack, Cursor) with multiple MCPs, skills, and plugins, a naive system simply passes raw JSON calls back and forth.

Without durable business context, evolutionary memory, and multi-modal source artifacts, AI runs blind tool calls and produces disconnected work ("work is no work"). Furthermore, treating embeddings or current scalar database values as the sole truth erases the chronological evolution of business decisions and destroys the audit trail.

Decision

  1. The Four Canonical Flows:

    • Flow A (Structured Business Data): Ingests structured business records (Accounts, Contacts, Deals, Tickets, Projects, Invoices, Tasks) from connected systems (Salesforce, HubSpot, Jira, Stripe, SAP, etc.). Represents the operational skeleton.
    • Flow B (Contextual Evidence & Artifacts): Ingests the unstructured material surrounding those records (emails, Slack threads, meeting transcripts, call recordings, PDFs, contracts, images, source files, attachments). Preserves the raw source artifacts and extracts structured metadata, entities, and evidence.
    • Flow C (Conversational & AI Cognitive Memory): Ingests the human–AI work sessions, ChatGPT/Claude conversations, Twin reasoning trails, AI-generated analyses, proposals, and decisions. Passes through Triage and staging before durable promotion to Organizational Memory.
    • Flow D (Platform Binding & Governance): Binds Flows A, B, and C to Identity, Tenant, Workspace, Principal, Role, Delegated Authority, ConnectionBindings, and Security Policies.

  2. RAG and Embeddings are Retrieval Mechanisms, Not the Spine:

    • Vector embeddings, semantic indexes, and RAG pipelines are cognitive retrieval mechanisms over the Spine Fabric, never substitutes for the underlying Spine or the raw artifacts.
    • Invariant: Embedding ≠ Source Artifact. The Spine preserves the original multi-modal files (PDFs, images, audio, markdown, documents) alongside their parsed chunks, embeddings, entity relationships, and cryptographic evidence hashes (evidence_refs).

  3. Preservation of Evolutionary Memory (Temporal Continuity):

    • The Spine is not a point-in-time scalar database that overwrites status = Healthy with status = At Risk.
    • The Spine explicitly preserves the temporal evolutionary continuum:
      $$\text{Initial State} \longrightarrow \text{Signals / Anomaly} \longrightarrow \text{Discussions} \longrightarrow \text{Decisions} \longrightarrow \text{Commitments} \longrightarrow \text{Actions} \longrightarrow \text{Reconciled Outcome}$$
    • Allows exact retrospective reconstruction of what happened, why it happened, what was promised, who approved it, and what changed over days, weeks, and months.

  4. Unlocking Universal Interaction Grammars:

    • Ask Your Spine: Traverses historical continuity, multi-modal evidence, and cross-system relationships to assemble what happened, why it mattered, what changed, and what remains open.
    • Ask Your Twin: Engages the digital coworker over authorized active context, current work objectives, and immediate evidence.
    • Assign Your Twin: Dispatches autonomous, bounded execution because the Twin possesses authorized business context, past decisions, available capabilities, and policy guardrails.
    • Reconcile to KPI: Every execution verifies whether the business metric (NRR, SLA, Health Score) moved in reality, closing the loop back into future context.

Alternatives Considered

  • Metadata-only Spine: Rejected. If the Spine only holds row IDs and tool names, the Twin has no situational awareness and cannot reconstruct meaning.
  • Pure Vector RAG without Canonical Entities: Rejected. Vector similarity cannot enforce deterministic business invariants, financial arithmetic, or role-based access control.
  • Destructive State Overwrites: Rejected. Overwriting past states destroys organizational memory and leaves agents unable to explain historical causality.

Consequences

  • The Ingestion Engine must route incoming streams through deterministic normalization (Flow A), artifact preservation & chunking (Flow B), and Triage staging (Flow C).
  • Storage architecture must explicitly maintain the separation between the Blob/Evidence Plane (R2/S3), Structured Truth (D1/Postgres), Vector Retrieval (Vectorize), and Cache/Projections (KV/Redis).
  • Every promoted memory or canonical mutation must carry provenance and evidence links back to the original source artifact or interaction transcript.