Knowledge-Action Infrastructure (KAI)

• KAI is an architectural pattern that operationalizes knowledge artifacts to directly influence and drive action.
• It integrates diverse representations—from ontologies to digital twins—into a unified, executable framework.
• The system emphasizes governance, provenance, and closed-loop adaptation to ensure trusted and reliable operations.

Knowledge-Action Infrastructure (KAI) is an architectural pattern in which knowledge is represented, maintained, governed, and operationally delivered so that it can directly influence or perform operational behavior. In recent work, the term appears both explicitly and through closely aligned formulations: an embedded world model that unifies prediction and intervention in embodied cognition, a discipline of “knowledge architecture” for executable organizational knowledge, a “living knowledge infrastructure” for agent skills, and an “agent-actionable operational scaffold” for scientific simulation (Wang, 2 Mar 2025, Garralda-Barrio, 1 Jul 2026, Huang et al., 15 Jun 2026, Li et al., 18 May 2026). The common premise is that the managed unit is not only data records or documents, but knowledge artifacts, world models, procedures, skill assets, and provenance-bearing representations that can be queried, reasoned over, validated, and acted upon. KAI therefore occupies the boundary between knowledge representation and operational systems: it is neither a passive repository nor a purely reactive controller, but a structured substrate coupling knowledge layers, action layers, and feedback loops.

1. Conceptual scope and historical lineages

A concise formulation of KAI appears in the AUKAI framework, where knowledge and action share “a common objective, common representations, and common feedback signals,” and where the infrastructure is decomposed into a knowledge layer, an action/control layer, and a monitoring, evaluation, and adaptation layer (Wang, 2 Mar 2025). A conceptually parallel formulation appears in work on enterprise AI systems, where “knowledge architecture” is defined as the discipline concerned with the architectural guarantees required for executable organizational knowledge, namely knowledge that is directly consumable by humans, models, agents, workflows, or applications to influence or perform operational behavior (Garralda-Barrio, 1 Jul 2026). Read together, these formulations characterize KAI as an infrastructure in which knowledge is operationalized rather than merely stored.

Several earlier and adjacent traditions supply the intellectual background. Direct-representation knowledgebases were proposed as an alternative to document-centered digital libraries, with realist semantic modeling of everyday activities, technologies, and infrastructures organized through a foundry of ontologies grounded in BFO (Allen et al., 2017). Networks of knowledge engines proposed a different operationalization, representing expert knowledge as automated algorithms that process information-only inputs and outputs, and composing them into executable workflows that generate requested information at runtime (Bergmair et al., 2018). More recent work on community-driven AI knowledge resources argues for interoperable knowledge modules, reasoning services, and open engineering frameworks so that software agents and robots can use curated, human-verifiable knowledge to plan, decide, act, and self-correct (Chaudhri et al., 19 Jun 2025). This suggests that KAI is best understood not as a single canonical system design, but as a family of architectures in which structured knowledge becomes an executable substrate for action.

The socio-technical reading of the concept is equally important. Knowledge infrastructures have been defined as “robust networks of people, artifacts, and institutions that generate, share, and maintain specific knowledge,” and work on science communication and aerospace engineering emphasizes that technical artifacts alone do not constitute infrastructure unless workflows, communities, and governance arrangements stabilize them over time (Wittenborg, 6 Mar 2026). KAI extends that perspective by making the action dimension explicit: knowledge artifacts are designed from the outset to support repeatable, traceable, and governed decisions, workflows, and interventions.

2. Architectural patterns and system layers

A recurring architectural pattern is a pipeline from sources to action. In knowledge-centric information systems, the reference model is: artifact sources, knowledge ingestion, knowledge repository, governance and provenance, knowledge serving, and consumers. The repository is commonly divided into raw, curated, and operational layers, where raw knowledge is captured with minimal transformation, curated knowledge is normalized, enriched, linked, and reviewed, and operational knowledge consists of consumer-ready knowledge views and bundles (Garralda-Barrio, 1 Jul 2026). This layering already encodes a transition from representation to action, because the operational layer is explicitly “ready to drive actions” by humans or agents.

Embodied and cyber-physical variants instantiate the same pattern with world-model structure. In AUKAI, raw observation $x_t$ is preprocessed to $\tilde{x}_t$, mapped into state representation $S_t = g_\phi(\tilde{x}_t)$, accumulated into memory $$h_t = f_{\text{memory}}(S_t, h_{t-1}; \Theta_{\text{mem}})$$, and used by a world model together with an intervention-utility function and policy (Wang, 2 Mar 2025). In KG-MAS, a centralized Knowledge Graph serves as a dynamic, shared world model for a Multi-Agent System, with a System Setup KG for static configuration and a System Data KG for dynamic operational state; agents query this KG for decision-making and update it with real-time state information (Abdela, 11 Oct 2025). In LLM-augmented semantic digital twins, the semantic layer is an RDF graph instantiating infrastructure, regulation, and event ontologies, and acts as the knowledge backbone over which reasoning, simulation, and optimization operate (Li et al., 9 Aug 2025).

A second recurring pattern is separation of write-side and read-side concerns. Knowledge architecture explicitly advocates a knowledge read/write split: authoring, curation, and approval on the write side versus retrieval and runtime consumption on the read side, each with different controls and consistency expectations (Garralda-Barrio, 1 Jul 2026). SkillWiki applies a similar separation to agent skills by decoupling long-term knowledge repositories from continuously updated, versioned, and governed skill repositories (Huang et al., 15 Jun 2026). In both cases, KAI is not merely a storage system; it is an operational fabric that distinguishes between how knowledge becomes authoritative and how it is consumed at execution time.

3. Knowledge representations and operational semantics

KAI is representation-plural rather than representation-monist. Knowledge architecture explicitly states that “Markdown bundles, schemas, ontologies, knowledge graphs, embeddings, vector indexes, relational tables, and object stores may all participate,” and treats representation as one layer of the architecture rather than the whole architecture (Garralda-Barrio, 1 Jul 2026). The foundry approach to human activities and infrastructures makes the same point from an ontological perspective: a broad domain requires not one ontology but a coordinated collection of ontologies and related resources, organized by shared upper-level commitments and linked through common relations (Allen et al., 2017). The implication is that a KAI must support coexistence and interoperability across symbolic, statistical, and document-centric forms.

A strong formalization of the operational layer is given by the Action Units framework. There, an action unit is represented as

$\text{AU} = (I, O, P, AC, G),$

where $I$ is the input unit, $O$ the output unit, $P$ the plan specification unit, $AC$ the applicability conditions unit, and $G$ the objective unit (Vogt, 2 May 2026). Three operation classes are distinguished: epistemic action units, transformational action units, and intervention action units. The framework also defines conditional action units as executable IF–THEN structures, so that knowledge graphs can function as graph-native decision-support systems. This is significant because it separates structural actionability from contextual applicability: a representation can support an operation in principle while still requiring context-sensitive validation before use.

Other representations specialize this principle for particular domains. ORKG models scholarly content through the minimal core entity ResearchContribution and represents fine-grained scholarly knowledge as statement-level graph structures rather than bibliographic metadata alone (Jaradeh et al., 2019). Aerospace.Wikibase models aerospace engineering around the question “Which Aerospace Engineering Process is completed by which Software using which Data, in which Format, and with which Schema?”, using OWL-aligned classes and Wikibase properties to encode process–software–data relations (Wittenborg et al., 5 Mar 2026). SkillWiki represents each skill as a governed capability asset with identity, classification, lifecycle state, interface specification, implementation, evaluation contract, provenance information, graph relations, and runtime metrics (Huang et al., 15 Jun 2026). Across these systems, KAI appears as a representational regime in which facts, procedures, capabilities, and evidential links are all first-class objects.

4. Action, control, and closed-loop adaptation

The defining feature of KAI is that it does not stop at representation. In AUKAI, the knowledge and action core is formalized through a predictive world model and intervention utility,

$\tilde{x}_t$0

with action chosen by

$\tilde{x}_t$1

and parameters optimized by the unified objective

$\tilde{x}_t$2

Multi-scale error feedback ties perception, memory, world modeling, and policy adaptation into a single closed loop (Wang, 2 Mar 2025). This is a canonical KAI formulation because prediction and intervention are co-optimized rather than separated into distinct subsystems.

Digital-twin variants operationalize a similar loop at the level of semantic reasoning and physical planning. In LSDTs, LLMs extract attribute snippets and constraint snippets from unstructured regulatory and technical documents, instantiate them as RDF triples and executable rules, reason over candidate layouts, update semantic state with environmental inputs according to $\tilde{x}_t$3, and feed the resulting knowledge into optimization and control (Li et al., 9 Aug 2025). The consequence is a planning system in which regulations, physical assets, and events are coupled inside one semantic substrate. In the Maryland offshore wind case study, this supports regulation-aware layout optimization and rule-driven state changes such as turbine shutdowns during Hurricane Sandy (Li et al., 9 Aug 2025).

Skill and simulation infrastructures make the control loop even more explicit. SkillWiki defines a lifecycle from Raw Experience through Candidate, Draft, Verified, Released, Degraded, Deprecated, and Archived, and monitors runtime health with

$\tilde{x}_t$4

Execution statistics, failure evidence, and reflection memories trigger maintenance proposals, version evolution, decomposition, repair, or deprecation through a Git-style governance workflow (Huang et al., 15 Jun 2026). KISS decomposes operational expertise into validated modelling operators, staged domain protocols, and diagnostic recovery mechanisms. In a 3,000-trial coupled-hydrology benchmark, KI-equipped agents produced physically plausible, verifiable end-to-end simulations in up to 84% of trials, whereas agents without KI plateaued below 40%; the same construction pattern was then packaged by the Knowledge Dissection Toolkit across 117 additional process-based models in 14 Earth-science domains (Li et al., 18 May 2026). These systems demonstrate that KAI is not only about choosing actions but about constraining execution, checking progress, and repairing failures.

5. Provenance, governance, and trust

Because KAI makes knowledge operational, provenance and governance become safety-critical rather than merely administrative. In knowledge-centric systems, governance determines who can create, modify, approve, access, publish, archive, or delete knowledge artifacts, and in agentic settings it also determines which knowledge can be activated for which task (Garralda-Barrio, 1 Jul 2026). Knowledge quality is correspondingly broader than conventional data quality: it includes freshness, consistency, grounding, contradiction management, completeness, and operational safety (Garralda-Barrio, 1 Jul 2026). A KAI without these controls may remain queryable, but it cannot be trusted to drive action.

Provenance is the second pillar. Knowledge provenance generalizes data lineage by requiring that a knowledge artifact be traceable to source artifacts, extraction and transformation steps, model-mediated synthesis, human review and approval, and runtime context assembly (Garralda-Barrio, 1 Jul 2026). SkillWiki implements this through a Skill Provenance Graph connecting knowledge sources, skills, executions, validations, versions, and inter-skill dependencies in a unified heterogeneous representation (Huang et al., 15 Jun 2026). ORKG similarly treats provenance and versioning as part of the persistence layer, enabling statement-level metadata and the evolution of curated scholarly knowledge (Jaradeh et al., 2019). The operational implication is that when an agent acts, a planner simulates, or a workflow executes, the system can expose what knowledge was used, where it came from, and how it was transformed.

The Action Units literature generalizes this requirement into the TripleA Principle: Actionability, Applicability, and Auditability (Vogt, 2 May 2026). Actionability requires that knowledge be structurally prepared for operations; applicability requires explicit conditions under which that knowledge is valid in a specific context; auditability requires that evidence, validation history, and conditions of validity remain traceable across applications. This principle is especially important in domains where knowledge can be executable yet not contextually justified. A plausible implication is that TripleA functions as a post-FAIR criterion for deciding whether a knowledge resource has crossed the threshold from being machine-readable to being operationally trustworthy.

6. Domain-specific instantiations

The range of instantiated KAIs is broad. In aerospace engineering, Aerospace.Wikibase.Cloud is an open, collaborative knowledge infrastructure built on Wikibase and initially populated with over 700 domain-specific terms extracted from a systematic literature review. The resulting knowledge graph contains 14,886 nodes and 37,897 triples, with a core domain model linking processes, software tools, data items, data format specifications, data models, and research contributions (Wittenborg et al., 5 Mar 2026). Its central organizing question—“Which Aerospace Engineering Process is completed by which Software using which Data, in which Format, and with which Schema?”—makes the action orientation explicit.

In scholarly communication, ORKG is a knowledge-graph-based infrastructure for acquiring scholarly knowledge in machine-actionable form. Its layered architecture separates persistence, domain model, interoperability, features, and services, while content-centric entities such as ResearchContribution, ResearchProblem, ResearchMethod, and ResearchResult support actions such as state-of-the-art comparison and similarity-based recommendation (Jaradeh et al., 2019). In public-health planning, a knowledge-based infrastructure for Quebec multi-stakeholder partnerships models objectives, strategies, actions, indicators, and data sources in OWL 2 so that epidemiological portraits can feed diagnosis, strategy formation, annual action plans, and evaluation (Addy et al., 2019). In industrial robotics, KG-MAS uses a centralized Knowledge Graph as a shared world model for autonomous agents that coordinate heterogeneous physical and digital assets through semantic descriptions and model-driven agent generation (Abdela, 11 Oct 2025).

Embodied and scientific instantiations push KAI toward autonomous operation. AUKAI casts embodied cognition as an embedded world model with utility-guided intervention, hybrid neural-symbolic modules, and multi-scale feedback (Wang, 2 Mar 2025). LSDTs couple extracted regulatory knowledge to semantic digital twins for offshore wind planning and extreme-event adaptation (Li et al., 9 Aug 2025). KISS externalizes scientific simulation expertise into operator, protocol, and diagnostic scaffolds for agentic Earth science (Li et al., 18 May 2026). In materials science and chemistry, the 2025 LLM hackathon explicitly separated “Knowledge Infrastructure” from “Action Systems,” but the reported systems increasingly combine retrieval, reasoning, tool use, structured memory, and laboratory or simulation execution into integrated workflows (Roy et al., 4 May 2026). Taken together, these cases show that KAI is not confined to a single technical stack; it is a recurring architectural response to domains where knowledge must be both semantically rich and operationally consequential.

7. Open problems and future directions

The main open problem is not whether KAI is possible, but how to make it reliable at scale. Knowledge architecture identifies unresolved questions in representation, synchronization, provenance, quality, governance, and evaluation, especially once the managed unit shifts from records to knowledge artifacts with semantic drift, authority changes, and operational consequences (Garralda-Barrio, 1 Jul 2026). The community-driven vision for a new AI knowledge resource adds that modern systems need expressive, verifiable modules, reasoning services, open engineering frameworks, contributor conventions, and social structures comparable to a public utility rather than a static database (Chaudhri et al., 19 Jun 2025). This suggests that KAI research is as much about institutional design as about ontology engineering or agent orchestration.

A second open problem concerns stability under continuous evolution. SkillWiki explicitly notes that long-term stability of evolving skill ecosystems remains untested, especially as execution-driven feedback modifies released skills over time (Huang et al., 15 Jun 2026). KISS shows that operational expertise can be extracted and packaged across many scientific models, but it also notes that such scaffolds do not solve structural model uncertainty or guarantee appropriate model choice for policy questions (Li et al., 18 May 2026). AUKAI’s convergence, stability, and near-optimality claims are conditional on assumptions such as sufficient exploration, bounded noise, appropriate learning rates, and manageable non-convexity; when those assumptions fail, convergence and stability may fail as well (Wang, 2 Mar 2025). KAI therefore inherits the familiar theoretical problem of maintaining guarantees in open, non-stationary settings.

A third open problem is socio-legal rather than algorithmic. Work on aerospace and science communication repeatedly emphasizes cultural resistance to sharing, intellectual-property constraints, safety regulation, and the mismatch between project time and infrastructure time (Wittenborg et al., 5 Mar 2026, Wittenborg, 6 Mar 2026). The broader implication is that KAI cannot be reduced to RAG, vector search, or agent frameworks. It requires durable governance, provenance-rich operational memory, explicit action models, and communities willing to maintain and contest them. Where those conditions hold, the literature suggests that KAI can function as executable organizational memory, embodied world model, governed skill ecosystem, or agent-actionable scientific commons; where they do not, the system remains a repository with automation attached rather than a genuine knowledge-action infrastructure.