cF Architecture

Highly isolated layers.

The cogniFactum architecture is built from the following highly isolated layers:

Build Phase

From activity (L3) to task (L4)

The customer’s processes are first captured at activity level (L3), based on cF’s standard process repository (cF’s Brain), and then enriched to task level (L4). The result is the rational Customer’s Brain 1.

Deterministic Guardrails

Level 4 – Customer’s Brain 1 Rational

The foundation of the architecture is a relational PostgreSQL database. It holds the customer’s operating model with all roles, phases, activities (L3) and atomic tasks (L4) according to the “4C standard” (Complete, Consistent, Compliant, Correct) – plus, for every task, references to the documents, templates, APIs and data sources it uses. The content of these objects is not stored here; it remains in the respective management systems (e.g. document management, API management). This layer acts as a hard guardrail that prevents AI agents from “inventing” processes (hallucinations).

Semantic Memory

Level 4 – Customer’s Brain 2 Analytical

The content of the documents, templates, APIs and data sources referenced in Customer’s Brain 1 is made available for AI analysis – either indexed in a vector database or retrieved directly from the source systems (e.g. document management, API management) via RAG. The source systems remain the owner of the content. The Large Language Model (LLM) uses this knowledge as a “cognitive analyst”.

The Deterministic Bridge

UUID Filtering

Every task in Customer’s Brain 1 carries a UUID (Universally Unique Identifier) that serves as a strict metadata filter. When retrieving knowledge, an AI agent may only access the content in Customer’s Brain 2 that is linked to the UUID of the task currently being executed.

Orchestration & Integration

BPMS, Agents & MCP

The processes defined in Customer’s Brain 1 are executed by a BPMS with robotics (e.g. Camunda, Pega), which also orchestrates AI agents within defined process steps. Where no BPMS is in place, or for agent work outside defined processes, a standalone multi-agent framework (such as LangGraph or n8n) can be connected. Communication with enterprise APIs runs centrally via the Model Context Protocol (MCP), which acts as a standardized translator and security hub.

Compliance & Read-Only Principle

Human-in-the-Loop

Generative AI models never receive direct write access to the deterministic process database (Customer’s Brain 1 – read-only). Instead, change proposals are generated that must be approved according to the “human-in-the-loop” principle. In addition, every system decision is logged in an audit-proof manner in an immutable State Management & TX log.

cogniFactum AI Platform Architecture

Build phase

cF’s Brain

L3 · 4C standard repository

Customer’s Brain L3

L3 · activities

Customer’s Brain L4

L4 · tasks + references

Two Brain concept

Customer’s Brain L4 Rational

Rational Brain · L4 · deterministic

PostgreSQL
Roles · phases · activities · tasks
References: documents, templates, APIs, data sources
4C standard

← read-only UUID filter · RAG →

LLM / AI agents

Analysis

↓ change proposal

Human-in-the-Loop

Approval before any change to Customer’s Brain 1

Customer’s Brain L4 Analytical

Analytical Brain · L4 · RAG

BPMN
Work instructions
Compliance rules
Regulations
Enterprise architecture
CMS documents
Standard processes of other industries
BIAN

Indexed in a vector database or retrieved live from the source systems

Execution & governance

Orchestration

BPMS & robotics (e.g. Camunda, Pega)
AI agents within defined process steps
Agent framework (optional, via MCP)

Integration

Model Context Protocol (MCP)
Enterprise APIs · API management
Applications
Manual processing
External sources

Governance & Audit

State Management & TX log
Immutable · audit-proof
Read-only for GenAI

cogniFactum AI Platform Architecture — Click to open in full size