Modern transformer-based AI models reconstruct meaning from context on every call. There is no separate reasoning module sitting on top of a data store. Context is the computational substrate, and the system's ability to maintain coherent meaning across it becomes the limiting factor when scaling complex AI workflows - especially across organizational silos.

Most users of AI think of context as chat history. In reality, context is the entire information environment that the model reasons over at inference time: prompts, retrieved documents, tool outputs, prior turns, scope rules, and working state. This environment is typically unstructured, and failures emerge at both the AI model level (lost-in-the-middle effects, context poisoning, compounding errors, semantically irrelevant retrieval) and the operational level (lost thread, erased history, intent drop, bloat, rot, and leakage). Only 23% of enterprise AI failures trace to model performance(Folio3, 2026). The remaining 77% come down to strategy, governance, and the absence of structured context.

Extending the context window does not solve this and only increases the surface area for meaning degradation. Addressing these bottlenecks requires two capabilities: semantically aware retrieval that ensures the right information enters the context and effective memory and state management that preserves coherent meaning across steps, actors, and time. This session presents how the semantic layer (taxonomy, ontology, knowledge graph, and business metadata) provides the foundation for the former, and how the context layer, structured as enduring, major, and minor context, addresses the latter. The session will provide real-world examples of emerging knowledge, data and AI architectures and practical guidance on how the semantic layer anchors the context layer and how to curate context across tiers in enterprise environments.

Speakers

Urmi Majumder
Principal Enterprise Architect
Enterprise Knowledge

Urmi Majumder is a Principal Architect at Enterprise Knowledge, where she leads system architecture, design, and implementation of a broad range of enterprise solutions. She has over 15 years of industry experience leading the development of technical solutions in support of a wide variety of federal and commercial clients by integrating open-source, SaaS, and COTS tools and establishing the connection between these tools and their business users. Her diverse portfolio includes the design and development of data-centric solutions that squarely sit at the intersection of KM, data, and AI.

Fernando Aguilar Islas
Senior AI Architect
Enterprise Knowledge

Fernando Aguilar is a Senior Solutions Consultant (Tech Lead and AI Engineer/Architect) at Enterprise Knowledge. He designs business-aligned semantic layers, knowledge graphs, and AI/ML solutions that make enterprise context machine-readable and governed. Most recently, he led work on graph analytics in the semantic layer as an architectural framework for knowledge intelligence and on unlocking knowledge intelligence from unstructured data. His delivery portfolio includes an enterprise identity graph and data catalog, AI-augmented content analysis, a graph-ML recommender for public safety, and generative-AI-assisted taxonomy for a global bank. He focuses on context engineering—hybrid retrieval (keyword, content vectors, graph vectors), unified entitlements, and 360° observability, to deliver grounded, transparent, and reusable enterprise AI.