Cognitive Architectures from A–Z.
Here you will find explanations of key Cognitive Architectures terms used frequently on this website — from architectural principles and system layers to governance and operations. Click on a term to view its definition — the collection is continuously growing and regularly updated to reflect new developments.
Missing evidence results in no answer rather than a wrong one — the system would rather remain silent than fabricate.
The handover of an AI-generated insight to an operational system, for example as a task in the CRM.
The component that translates insights from a multi-agent system into concrete actions in target systems.
A specialised AI component with a clearly defined task — such as analysis, validation or planning.
A security principle whereby only explicitly permitted actions or tools are accessible — everything else is blocked by default.
The control layer that ensures traceability and audit trails for every AI decision — including validation, monitoring and audit.
The property whereby all decisions and processes of a system are traceable and verifiable.
The property of a system to make every statement fully traceable after the fact.
A system for the curated, traceable generation and maintenance of ontologies — including governance and a review process.
The system may summarise information, but must not extrapolate beyond the available evidence.
Continuous learning of a system from real-world outcomes and feedback loops — not just from training data.
Multi-layered AI systems in which every component carries a clear responsibility — from data and governance to operations. As a portfolio brand, the term refers to M2’s entire offering; as a system type (Cognitive Architecture, singular), it refers to the individual architectural approach.
Persistent storage for validated insights that a system can reuse without having to derive them again.
The operational layer of an AI system that brings together semantics, agents, actions and governance, turning findings into concrete actions — not just insights.
The central control unit of a multi-agent system: it decides which agent is used, which tool is deployed and what evidence is required.
The control layer for identity, policies, routing and security decisions within a Cognitive Architecture.
Agents may act independently, but only within clearly defined roles, rules and boundaries — autonomy is explicitly modelled, not implicitly permitted.
The same input and the same data lead to the same result and the same processing path — the counter-model to the randomness of generative systems.
Also known as evidence-first: all answers are based exclusively on verifiable sources and data and must never be freely “invented”.
The layer where the actual processing takes place — analysis, tool use, generation.
The validation of statements against real data sources, such as databases or documents, before an answer is returned.
The ability for humans to intervene in or review a system’s decisions at any time.
The layer that manages knowledge: documents, semantic models, memory and provenance.
Each agent is granted only the minimum permissions required to perform its task.
A distributed AI system consisting of multiple specialised agents, tools, data sources and control mechanisms that perform tasks collaboratively — not an extension of a chatbot, but a dedicated execution and decision layer.
A system design with multiple parallel processing paths — for example generative versus deterministic — instead of a single path for all requests.
The ability to fully observe a system’s behaviour — through logs, metrics and traces.
A structured description of business objects and their relationships to one another.
The transformation of insights into concrete operational measures — the transition from mere insights to action within the organisation.
Evidence of where data or statements originate and how reliable they are.
The retrieval of relevant information from data sources to support answers — often as part of a RAG approach.
The targeted restriction of AI degrees of freedom where predictability matters more than creativity — critical steps run rule-based rather than generatively.
An abstraction layer that translates technical data into business meaning.
The set of rules that defines which tools may be used and under what conditions.
The authority that checks every answer from a multi-agent system against its source before it is returned.
The checking of results before they are returned — often by an agent separate from the one generating the answer.
Missing a term, or does a definition not fit your use case? We are continuously expanding the glossary.
Back to Cognitive Architectures →
Secure Multi-Agent Systems of the Future.
Multi-agent systems promise greater reach — but also more states, more handovers and a larger attack surface. Our whitepaper shows how controlled autonomy, evidence-bound answers and resilient operations turn a risky idea into a productive, auditable system — illustrated by the real-world example of Cognitive Analytic Agents.
Download whitepaper