Context Graphs¶

Platforms: claude openai gemini m365-copilot

What Context Graphs Are¶

A context graph is a structured system that captures not just what information an AI agent uses, but why decisions were made and how facts relate to each other.

Think of it this way: a document gives an AI agent information. A context graph gives it understanding — the connections between entities, the reasoning behind decisions, and the temporal sequence of events that led to the current state.

Context graphs build on the concept of a knowledge graph (a network of entities connected by labeled relationships) but go further by encoding decision logic, confidence levels, and causal chains. In a knowledge graph, you might store that "Customer A bought Product B." In a context graph, you also capture why — the sales signal that triggered outreach, the objection that was overcome, and the precedent from a similar deal that informed the approach.

How Context Graphs Differ¶

Approach	What It Does	What It Misses
Context window	Holds the text (measured in tokens — chunks of text the model processes) visible to the model in a single conversation	No structure, no persistence, limited size
RAG	Retrieval-Augmented Generation — retrieves relevant documents from a database and injects them into the prompt	Finds related content but doesn't capture relationships between concepts
Knowledge graph	Maps entities and their relationships in a structured network of nodes (things) and edges (connections)	Static structure, no decision reasoning or temporal context
Context graph	Captures entities, relationships, decisions, reasoning chains, and temporal context in a queryable structure	Emerging technology, more complex to build and maintain

Each approach in this table builds on the ones above it. RAG helps an AI find relevant documents. A knowledge graph helps it understand how entities relate. A context graph helps it understand why things happened and what led to decisions — which is what agents need for multi-step reasoning.

Key Concepts¶

Nodes¶

Nodes are the things in a context graph. Unlike a traditional knowledge graph where nodes are mostly entities (people, companies, products), context graph nodes include:

Node Type	What It Represents	Example
Entity	A person, organization, product, or concept	"Acme Corp", "Q4 Revenue Report"
Decision	A choice that was made, with reasoning	"Chose vendor B because of compliance requirements"
Signal	An event or data point that triggered action	"Customer satisfaction dropped below 80%"
State	A snapshot of conditions at a point in time	"Pipeline status as of January 2026"

Edges¶

Edges are the connections between nodes — they describe how things relate. Context graphs use richer edge types than traditional knowledge graphs:

Edge Type	What It Captures	Example
Causation	One thing led to another	Signal "churn risk detected" → Decision "escalate to account manager"
Dependency	One thing requires another	Task "generate report" depends on State "data refresh complete"
Precedent	A past decision that informs a current one	Decision "pricing for Enterprise tier" references Decision "pricing for Mid-Market tier"
Temporal	Sequence and timing of events	Signal A occurred before Signal B, 3 days apart
Confidence	How certain a relationship is	Entity "likely competitor" connected with 0.7 confidence

Relationships¶

The combination of nodes and edges creates relationship patterns that agents can traverse:

Temporal chains — "What sequence of events led to this outcome?"
Conditional logic — "Under what conditions was this decision made?"
Confidence-weighted paths — "What's the most reliable chain of reasoning?"

Why They Matter for Agentic AI¶

Context graphs address specific limitations that surface when AI agents handle multi-step, real-world workflows:

Multi-step reasoning — Agents can trace chains of causation and dependency rather than relying on whatever fits in the context window
Structured memory — Decisions and their rationale persist across conversations and sessions, giving agents institutional knowledge
Auditability — Every recommendation can be traced back through the graph to the signals and precedents that informed it
Compounding organizational knowledge — Each interaction adds to the graph, making the system more capable over time rather than starting fresh each session

The Context Engineering Shift¶

The AI industry is undergoing a shift from prompt engineering (optimizing how you ask) to context engineering (optimizing what information the model sees).

The core insight: a well-structured, relevant context makes even a simple prompt produce excellent results. A perfectly crafted prompt with poor context still produces poor results.

This shift has practical implications:

Where you invest time changes — less on prompt syntax, more on building and curating the information your AI workflows consume
What you build changes — systems that assemble, filter, and structure context become more valuable than prompt template libraries
How you measure quality changes — success depends on whether the right context reached the model, not just whether the prompt was well-written

Context graphs represent the most structured end of this spectrum — purpose-built systems for assembling exactly the right context for each agent action.

Tools and Frameworks¶

Tool	What It Does	Best For
TrustGraph	Extracts knowledge graphs from documents with AI, supports context graph queries	Document-heavy workflows needing structured extraction
Graphiti	Temporal knowledge graph library for building agent memory	Agents that need persistent, evolving memory across sessions
LangGraph	Framework for building stateful, multi-agent workflows with graph-based orchestration	Complex agent orchestration with branching logic
Neo4j	Graph database for storing and querying relationship-rich data	Enterprise-scale knowledge and context graph storage
MCP	Model Context Protocol — a standard for connecting AI models to external data sources and tools	Connecting agents to live context from APIs, databases, and services

Relationship to Other Blocks¶

Context graphs intersect with every other building block:

Prompts — Context graphs provide the structured background that makes prompts effective. Instead of cramming context into a prompt, agents query the graph for exactly what's relevant.
Projects — Projects organize persistent context. Context graphs add structure within that context — not just "here are the files" but "here's how they relate."
Skills — Skills can use context graphs to make decisions based on precedent and historical patterns rather than just the instructions they contain.
Agents — Agents are the primary consumers of context graphs. Multi-step workflows benefit most from structured reasoning and memory.
MCP — MCP servers can expose context graph data as tools and resources, giving agents structured access to organizational knowledge.