Knowledge-Defined Architecture

• Knowledge-Defined Architecture is a paradigm that externalizes structured knowledge from design artifacts and expert insights.
• It enables systems to retrieve, reason over, and update explicit knowledge bases—enhancing decision accuracy and adaptability.
• The approach supports multi-domain applications by integrating deep knowledge bases, provenance layers, and context units for robust governance.

Knowledge-Defined Architecture is an architectural paradigm in which explicit knowledge becomes a primary organizing and operational substrate of the system rather than remaining implicit in prompts, static rules, raw telemetry, or neural parametric memory. Across software architecture automation, hybrid AI, networking, enterprise information systems, and cognitive architectures, the recurring design move is to externalize knowledge into structured artifacts such as deep knowledge bases, knowledge graphs, provenance layers, context units, or governed knowledge artifacts, and to make inference, control, generation, or execution depend on retrieving, reasoning over, updating, and validating those structures (Zhang et al., 26 Mar 2025, Singer et al., 2023, Garralda-Barrio, 1 Jul 2026).

1. Conceptual foundations

In the software-architecture literature, the term is stated directly: MAAD treats architecture as a form of knowledge-defined architecture, meaning that architecture is not generated from a single prompt but is constructed by integrating knowledge from existing system designs, authoritative literature, and architecture experts (Zhang et al., 26 Mar 2025). In enterprise information systems, the same shift is formulated as a transition from managing records and tables to managing knowledge artifacts, with the formal claim that when heterogeneous organizational knowledge is consumed operationally by humans, models, agents, workflows, or applications, information systems require architectural guarantees for knowledge artifacts analogous to, but semantically broader than, the guarantees that data architecture provides for records and tables (Garralda-Barrio, 1 Jul 2026).

A related formulation appears in hybrid AI. Thrill-K is presented as a knowledge-defined / knowledge-based hybrid architecture in which intelligence is stratified across instantaneous knowledge, standby knowledge, and external knowledge. This is explicitly contrasted with Class 1 end-to-end systems and with Class 2 retrieval-augmented systems whose richer relationships remain largely in neural embeddings; Thrill-K is a Class 3 architecture with a deep knowledge base containing explicit structure (Singer et al., 2023). In the networking literature, KCN defines knowledge as “processed information that creates awareness of the surrounding environment and enables the network to react automatically and timely to emerging conditions,” then moves that knowledge to the center of the networking stack (Charalambides et al., 2020).

These formulations converge on a common thesis. Knowledge-defined architecture does not merely add retrieval or metadata to an existing system. It redefines what the architecture is centered on: not only data movement, not only model inference, and not only control logic, but a maintained body of structured, updateable, and operationally consumable knowledge. This suggests that the term names a family of architectures united by a shift from implicit competence to explicit knowledge mediation.

2. Recurring architectural patterns and managed units

A recurrent property of knowledge-defined architectures is stratification. Thrill-K formalizes 3LK as instantaneous knowledge, standby knowledge, and retrieved external knowledge, with the default flow “Neural network first,” then consultation of the deep knowledge base through reasoned extraction, and only then external retrieval if needed (Singer et al., 2023). Enterprise knowledge architecture presents an analogous reference model: Artifact Sources → Knowledge Ingestion → Knowledge Repository → Governance and Provenance → Knowledge Serving → Consumers, together with the raw–curated–operational layering of knowledge (Garralda-Barrio, 1 Jul 2026). In 6G management, the layered form becomes Data Plane → Digital Twin → Knowledge Plane → Management Plane → Control Plane → Data Plane, yielding a closed-loop autonomic control system (Bilen et al., 27 Sep 2025). Condense expresses the same idea in a communication setting through Atomic Function Computation layer, Network Function Computation layer, and Application layer, so that the network computes the required function instead of merely forwarding raw data (Vukobratovic et al., 2016).

A second recurrent property is the introduction of an explicit unit of management. The managed unit is no longer only a packet, record, prompt, or tensor.

Context	Managed unit	Architectural role
KCN	knowlets	local knowledge atoms for local control (Charalambides et al., 2020)
Knowledge architecture	knowledge artifact	persistent item contributing to organizational understanding or execution (Garralda-Barrio, 1 Jul 2026)
Context Kubernetes	context unit $u = (c, \tau, m, v, e, \pi)$	smallest addressable piece of organizational knowledge (Mouzouni, 13 Apr 2026)
Knowledge Activation	Atomic Knowledge Unit (AKU)	minimal self-contained bundle for one coherent action (Bakal, 16 Mar 2026)
Companion	microtheory / provenance event	logical context and epistemic trace of knowledge (Nakos et al., 2024)

The same architectural logic appears in synthesis systems. The knowledge-driven AutoML architecture unifies pipeline synthesis and deep feature synthesis around a Knowledge System (KS) comprising a knowledge base (KB) and a constraint solver (CS); the control flow repeatedly queries KS, receives valid operation nodes, executes partial solutions, and uses runtime outputs to decide subsequent steps (Cofaru et al., 2023). ModuleNet constructs a knowledge base from modules extracted from pretrained CNNs and searches over the macro-space $|\Omega| = n^c$, where each position selects one inherited module (Chen et al., 2020). TRADE treats architecture search under knowledge distillation as a distillation-aware optimization problem and uses trust-region Bayesian optimization over a search space containing MBConv, Fused-MBConv, and Transformer blocks (Kim et al., 2022). This suggests that knowledge-defined architecture can denote either a runtime system centered on explicit knowledge or a design-time synthesis process whose search space is itself defined by reusable knowledge.

3. Software architecture and architecture design automation

In software architecture design, knowledge-defined architecture is articulated most directly by MAAD. The framework simulates four human roles—Analyst, Modeler, Designer, and Evaluator—to transform an input Software Requirements Specification into architecture artifacts. The Analyst extracts requirements and architecture-significant requirements, the Modeler selects technologies and architectural styles and constructs logical and physical conceptual views, the Designer produces implementation-oriented blueprints, and the Evaluator checks alignment with the SRS, diagnoses mismatches, and sends refinement suggestions back to the relevant agents (Zhang et al., 26 Mar 2025). The workflow is explicitly collaborative and iterative rather than one-shot; the process ends only when the Evaluator confirms alignment with the requirements. Its knowledge sources are equally explicit: existing system designs, authoritative literature, and architecture experts.

The later MAAD study operationalizes that vision in greater detail. It incorporates Retrieval-Augmented Generation (RAG) over architectural standards and texts, including ISO/IEC/IEEE 42010:2022, ISO/IEC/IEEE 42020:2019, Software Architecture in Practice, Clean Architecture, and Fundamentals of Software Architecture. It also introduces a hierarchical memory with working memory, episodic memory, and semantic memory, and generates a 4+1 view model with 11 UML diagrams in PlantUML syntax, together with production-ready architecture documentation, an ATAM-based architecture evaluation report, and a requirements-architecture mismatch analysis report (Li et al., 31 May 2026). In quantitative evaluation across 10 case studies, MAAD outperformed MetaGPT on reported architecture-level metrics, with CD 1.77 vs 1.87, Coh 0.69 vs 0.36, IC 2.08 vs 0.00, and SC 0.03 vs 0.17. The same paper also states that architecture quality depends heavily on the underlying LLM’s reasoning capacity: GPT-5.2 and Qwen3.5 outperformed DeepSeek-R1 and Llama3.3 across most evaluation settings.

A complementary vision appears in architectural knowledge management. “From Scattered to Structured” proposes an automated pipeline that extracts architectural knowledge from heterogeneous artifacts, links them through traceability, detects and resolves inconsistencies, and consolidates the result into a structured knowledge base supporting architecture conformance checking, change impact analysis, and natural-language question answering via RAG (Keim et al., 27 Jan 2026). Here the architecture is not only something produced by knowledge; it is also something maintained as structured knowledge over time. That emphasis on conformance, inconsistency detection, and erosion control is a distinct but compatible interpretation of knowledge-defined architecture.

4. Networking, cyber-physical control, and semantic management

In networking, the paradigm is often expressed through Knowledge-Centric Networking (KCN) or Knowledge-Defined Networking (KDN). KCN places knowledge at the center of the networking landscape and organizes it hierarchically into local knowlets, domain-level knowledge, and inter-domain exchange. It distinguishes facts from advanced knowledge, and it supports three exchange styles: Knowledge as a Tool (KaaT) within a domain, Knowledge as a Service (KaaS) across stakeholders under SLAs, and Knowledge as a Cloud (KaaC) through publish/subscribe-style sharing, with ICN and CBIS identified as enabling mechanisms (Charalambides et al., 2020). The defining shift is from packet transport and centralized management toward distributed knowledge creation, distribution, storage, and processing.

The 6G literature makes the control loop more explicit. The twin-assisted KDN architecture uses a five-layer stack—Data Plane, Digital Twin Plane, Knowledge Plane, Management Plane, and Control Plane—in which telemetry is normalized and forwarded to a digital twin, a GRU-based forecaster predicts short-term state transitions, a dynamic knowledge graph $$\mathcal{G}_{t+\Delta} = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$ is built from current and forecasted state, and a management layer performs zero-shot policy matching by semantic embedding and cosine similarity (Bilen et al., 27 Sep 2025). The paper reports sub-second response times even for novel scenarios, notable improvements in policy response time, SLA compliance rate, and intent matching accuracy, and low decision overhead because the forecaster is lightweight and policy matching requires no online retraining.

A related semantic-aware 6G framework treats semantics as a network-wide control variable rather than an endpoint codec property. Its Semantic Reasoning Module computes relevance by fusing task similarity, knowledge-graph context, and network urgency:

$$R(m) = \alpha\, S(\mathbf{s}(m),\mathcal{T}) + \beta\, C(m) + \gamma\, U(m), \quad \alpha + \beta + \gamma = 1$$

with the reported setting $(\alpha,\beta,\gamma) = (0.4, 0.3, 0.3)$ (Bilen et al., 13 Mar 2026). Routing and distortion control are then driven by semantic relevance and observed distortion. In ns-3, the reported gains are semantic delivery success improves by 12%, semantic distortion decreases by 22%, re-routing events drop by 44%, and throughput efficiency rises by 14% compared to shortest-path, load-based, and distortion-only baselines.

Industrial IT/OT autonomy adopts the same architecture in graph form. The proposed framework has Telemetry Collector, Knowledge Builder, Decision Engine, and Control Enforcer, with the system state represented as $$\mathcal{G}_t = (\mathcal{V}, \mathcal{E}, \mathbf{X})$$ and action selection formulated as utility maximization over a finite action set $\mathcal{A}$ (Bilen et al., 27 Sep 2025). The reported evaluation uses 50 interconnected nodes, 1000 time steps, and state updates every 10 ms, and it reports significantly lower decision latency, lower jitter and throughput variance, and higher stability than threshold-based, heuristic, and reinforcement-learning baselines. Condense pushes the same principle closer to the communication substrate: the network itself computes functions over IoT data via AFC and NFC layers, so that the destination receives the function evaluation needed by the application rather than the entire raw stream (Vukobratovic et al., 2016).

5. Enterprise knowledge infrastructure, governance, and cognitive architectures

Enterprise information systems reinterpret the paradigm as knowledge architecture. The central claim is that organizational knowledge is becoming executable infrastructure: something directly consumable by humans, models, agents, workflows, or applications to influence or perform operational behavior. The paper formalizes Definition 1 (Executable organizational knowledge) and Definition 2 (Knowledge Architecture), then maps familiar data-engineering guarantees onto knowledge-level guarantees: ETL → knowledge ingestion, CDC → knowledge change detection, lineage → provenance, catalogs → knowledge catalogs, materialized views → knowledge views, and medallion architectures → raw–curated–operational knowledge layers (Garralda-Barrio, 1 Jul 2026). It also states that LLM Wiki and the Open Knowledge Format (OKF) are early evidence of the transition, not the endpoint, and explicitly argues that RAG is only a serving technique, not a full architecture.

Context Kubernetes makes governance a first-class architectural concern. It defines six abstractions—Context Unit, Context Domain, Context Store, Context Endpoint, Context Runtime Interface (CxRI), and Context Operator—and expresses the desired state of enterprise knowledge through a YAML manifest for knowledge-architecture-as-code (Mouzouni, 13 Apr 2026). Its reconciliation loop monitors drift such as source_disconnected, context_stale, operator_unhealthy, anomaly, reliability_drift, and permission_change. The access-control core is the invariant

$P_{a_u} \subset P_u$

meaning agent authority is always a strict subset of human authority, together with Strong Approval Isolation for Tier 3 actions. Reported experiments include 26.5% of queries leaking cross-domain data in the ungoverned condition, staleness detected in under 1 ms with reconciliation, and 5/5 attacks blocked by the three-tier permission model.

Knowledge Activation focuses on institutional knowledge for agentic software development. It defines the Atomic Knowledge Unit (AKU) as the minimal self-contained bundle of intent, procedural knowledge, tool bindings, organizational metadata, governance constraints, continuation paths, and validators needed for one coherent action (Bakal, 16 Mar 2026). The architecture is justified by the Context Window Economy and by a formal notion of knowledge density,

$\rho(k, \tau) = \frac{v(k,\tau)}{c(k)}$

where $v(k,\tau)$ is task value and $|\Omega| = n^c$0 is token cost. Knowledge Activation is then the three-stage process of codification, compression, and injection, with agents traversing a graph of AKUs rather than loading monolithic documentation. This is a specifically enterprise form of knowledge-defined architecture: governance and execution-readiness are embedded in the knowledge primitive itself.

The cognitive-architecture tradition contributes a more formal knowledge stack. Companion treats knowledge as “the fuel for cognitive architectures” and organizes it across a physical layer, a logical layer based on microtheories, and an epistemic layer organized by provenance events (Nakos et al., 2024). Its representation stack includes Cycl, KRF files, NextKB, PlanB, and AllegroCache; the paper reports a knowledge base of over 700,000 facts, across 1,300 microtheories, in 1,000 KRF files. The addition of a provenance cache and the Archivist agent illustrates a recurring knowledge-defined requirement: explicit management of source, timestamp, update, and trust. At a methodological level, the knowledge-engineering reference-architecture paper generalizes the same idea into a six-step roadmap in which user needs → requirements → workflows → boxology patterns → reference architecture → software instantiation (Allen et al., 2024).

6. Limitations, misconceptions, and research directions

A recurrent misconception is that knowledge-defined architecture is equivalent to attaching retrieval to an otherwise unchanged model or controller. Several papers explicitly reject that simplification. Thrill-K distinguishes Class 2 retrieval-augmented systems from Class 3 systems with a deep knowledge base and reasoned extraction (Singer et al., 2023). The enterprise knowledge-architecture paper states that LLM Wiki and OKF are evidence of the shift, not the solution itself, and that RAG is only a serving technique (Garralda-Barrio, 1 Jul 2026). MAAD likewise argues that one-shot prompting produces hallucinated, inconsistent, or incomplete architectural decisions and replaces it with specialized agents, explicit knowledge sources, and evaluator-driven refinement (Zhang et al., 26 Mar 2025). This suggests that the distinctive property is not retrieval alone, but the architectural centrality of explicit knowledge, provenance, and validation.

The limitations are correspondingly broad. MAAD is explicit that explainability, trustworthiness, agent coordination, communication protocols, and adaptive workflows remain open problems, and its original paper does not provide a formal algorithm or detailed pseudo-code for full coordination (Zhang et al., 26 Mar 2025). Thrill-K identifies higher architectural complexity, new training algorithms, inconsistency handling, and knowledge-acquisition difficulty (Singer et al., 2023). Context Kubernetes states that its invariants are empirically enforced in a prototype rather than formally proven, and names tools such as TLA+ or Alloy as future verification directions (Mouzouni, 13 Apr 2026). Companion identifies provenance, deletion tracking, trust, distribution, and autonomous self-directed learning as unfinished parts of the knowledge stack (Nakos et al., 2024). Software-architecture knowledge management adds further issues: heterogeneous and multimodal artifacts, LLM hallucinations, distinguishing extraction errors from real inconsistencies, temporal validity, and security for sensitive materials (Keim et al., 27 Jan 2026).

Across these domains, the research trajectory is clear even when the implementations differ. Better knowledge injection, improved multi-agent coordination, stronger provenance and governance, more explicit semantic abstractions, and architectures that preserve freshness, traceability, and fitness for action are repeatedly identified as the next steps (Li et al., 31 May 2026, Bilen et al., 27 Sep 2025, Garralda-Barrio, 1 Jul 2026). Knowledge-defined architecture therefore appears less as a single canonical stack than as a general architectural response to a common problem: once systems must reason, act, or design with organizational, domain, or network context, that context must be represented, governed, and operationally delivered as knowledge rather than left implicit.