Context layer glossary

Accumulated cost from shortcuts in semantic infrastructure — grows like financial debt.

Memory of specific past events and interactions.

Memory of how to perform specific tasks and workflows.

Memory of general facts, definitions, and concepts.

The process of building the context window on the fly for each specific request.

Conflicting information from multiple sources present simultaneously in the context window.

Performance degradation caused by too many tools or instructions competing for the model's attention.

Designing the entire information environment in which an AI model operates.

The architectural system that curates and delivers context to AI agents — composed of Semantics, Operational State, and Provenance.

Keeping the context window current and relevant through summarisation, trimming, offloading, and ordering.

Pre-assembled, reusable bundles of context built for specific workflows.

Wrong or manipulated information entering the context and being treated as fact.

Gradual degradation as the context window fills with low-signal content over a long session.

The AI model's working memory — everything it can see right now.

Embedding decision intent, constraints, and success metrics into agentic systems as explicit context.

RAG enhanced with knowledge graph traversal — retrieves relationships between facts, not just similar text.

Independent AI agents that monitor and can block other agents' behaviour at machine speed.

Human approval required before each agent action.

Human sets intent upfront; agent acts autonomously; human monitors outcomes.

A database of entities and their relationships — enabling agents to reason across connected facts.

Model Context Protocol — a universal standard (USB-C for AI) for connecting models to external tools and data sources.

Everything that is currently true — world state, execution state, environment state.

Crafting better instructions and questions to improve model outputs.

The complete recorded history of where data came from, what decisions were made, and what actions were taken.

Retrieve-then-generate — look up relevant documents before answering, rather than relying on training data alone.