Architecture

How Cogniseam is built

The design of the memory, the graph, retrieval, and governance — and the guarantees the architecture makes. Everything on this page is shipped and tested, not roadmap.

The canonical architecture — every stage, one system
  1. Enterprise SystemsSlack · Microsoft 365 · Drive · ERP
  2. Connectorssync · live capture · CDC
  3. Normalizationone idempotent pipeline
  4. Knowledge Extractionfacts · entities · policies · skills
  5. Enterprise Memoryretrieval · history · supersession
  6. Knowledge Graphentities · relationships
  7. GovernanceRBAC · audit · approvals · encryption
  8. Executionanswers · skills · workflows · agents · API
  9. Business Outcomesfaster decisions · recovered revenue · provable compliance

Every section below is a zoom into one segment of this same flow — there is no second architecture diagram.

195
unit tests
ACLs, temporal supersession, CDC, SAML/SCIM, entitlements
51
end-to-end tests
auth/RBAC, SSO + SAML + SCIM against mock IdPs, billing
3
AI providers
Anthropic · OpenAI · Gemini behind one interface — switchable at runtime, BYOK
0
native dependencies
one Node process, one data directory — npm run setup on a fresh machine
~60
API routes
one Next.js deployable serves the UI and every endpoint
How knowledge flows

Four stages, one governed pipeline

  1. Ingest

    Knowledge arrives from Microsoft 365, Google Workspace, Slack, Notion, dev tools, uploads, and live meetings — normalized into one pipeline. Re-syncs are idempotent: unchanged content is skipped, changed content replaced in place.

  2. Structure

    Documents become atomic facts (categorized, confidence-scored, time-stamped), executable skills (trigger → steps → guardrails → owner), and typed knowledge-graph relations.

  3. Govern

    Access rules are stamped onto every chunk, fact, and skill at ingest. Retrieval filters by the caller's clearance before ranking — restricted content never reaches a model.

  4. Serve

    Cited answers for people; an authenticated Agent API, an MCP server, and portable SKILL.md export for agents; workflows that act under human approval gates.

The architecture

Six subsystems, one design

What an architect actually evaluates — how memory, structure, retrieval, and the agent surfaces fit together, without the parts list.

Memory
A temporal store, encrypted at rest. Changed knowledge is superseded with validity windows, never destroyed — “what was true last quarter, and when did it change?” is a query. Every fact carries provenance back to its source document and a freshness clock.
Knowledge graph
Typed relations — owns, depends on, approves, reports to — extracted alongside facts, forming the structural map of the organization. Edges are temporal like facts: the graph is point-in-time queryable, and every edge traces to the document that taught it.
Retrieval
Lexical and semantic ranking blended, with access control applied as a pre-filter before scoring. Works with zero AI keys and degrades gracefully: a provider outage returns ranked passages, never a failure — callers don't branch on AI availability.
Connectors
Every source flows through one normalize-then-ingest seam with change-data-capture semantics — stable external identity plus content hashing means syncing is repeatable without duplicates. Live meeting capture streams transcripts into the same pipeline.
Agent surfaces (API · MCP · SKILL.md)
Agents are first-class callers, not an afterthought: a key-authenticated Agent API returns answers, citations, and skills; an MCP server plugs the memory into any MCP client; skills export in the open SKILL.md format any agent can execute. Model-neutral by design — Anthropic, OpenAI, and Gemini behind one interface, bring your own key.
Deployment
One process, one data directory, no native dependencies — a single artifact that runs as SaaS today and is structurally ready for VPC or self-hosted deployment in regulated environments. Encryption, backups, and the audit trail travel with it.
Security & governance

Guarantees the architecture makes

Not policies bolted on afterwards — properties of how the system is built. The full evidence-backed posture lives in the Trust Center.

Permissions are a pre-filter

Access control runs before ranking, not after generation. Content a caller isn't cleared for is never scored, never retrieved, and never placed in a model's context window — the model can't leak what it never received.

Admins don't bypass the ACL

Private-group content is invisible even to workspace admins. Emergency access exists — as an owner-granted, time-boxed steward grant that writes to the audit log on grant, activation, and every read.

Agents are least-privilege callers

Each agent key maps to explicit audience scopes; the caller's identity is resolved server-side from the key, never from anything the agent claims. Workflows retrieve as the human who created the job.

Actions need a human

Delivery workflows pause at an explicit approve/reject gate and only ever open draft pull requests. A malformed approval defaults to rejection, never to action.

Knowledge carries its freshness

Every fact and skill has a decay-based review date; stale items surface in a queue, and new knowledge is checked against old — genuine contradictions are flagged with severity instead of being served silently.

History is never destroyed

Changed documents supersede their old facts and graph edges with validity stamps. The past stays queryable — for audit, for context, for “why did we decide this?”

Scalability

Today, and at ten thousand users

The honest version: what's validated now, and the documented seam each piece scales through — designed in from the start, not discovered later.

Validated todayThe documented seam
File-backed store, single writer, scale-to-zero idle costPostgres + pgvector behind the same retrieval interface — nothing above it changes
Vector similarity computed in-processSimilarity pushed into the database index
Pull-based connector sync, idempotent by designWebhook and streaming sync workers on the same normalize→ingest seam
Single region, single replicaHigh-availability replicas with service-level monitoring
Design principles

Five rules the architecture keeps

Not slogans — properties of how the system is built. Each maps to a stage of the flow above.

01
Memory before prompting.
Ground the model in governed memory, not a clever prompt.
02
Access before retrieval.
Clearance is checked before anything is ranked — never after.
03
Evidence before answers.
Every answer is cited to its source, or it says it doesn't know.
04
Governance before action.
Actions that leave a mark wait for a human.
05
Audit before automation.
Nothing runs unattended until it's provably on the record.
The outcomes

Governed memory, delivered as business value.

Trusted AIBetter decisionsInstitutional knowledgeFaster onboardingProvable complianceOperational excellence

Built agent-first

An authenticated Agent API, an MCP server for any MCP client, and skills that export as portable SKILL.md — so every agent your company runs consults the brain before it acts.