Skip to main content Context Graphs

• Progression
• Anatomy
• In practice
• Builder
• Craft
• Start

A working primer

Context graphs are how you ship reasoning , not just retrieval .

Most of what an organization knows doesn’t live in any one document; it lives in the connections between things:

• Who works with whom
• What depends on what
• Which decision led to which outcome

A context graph is how you write those connections down in a form a machine can actually use.

Start the primer Skip to the builder Written for Engineers & PD Reading time ~15 min Take away Exportable YAML / JSON / Cypher 01 The progression

Eight steps from language models alone to a graph that learns from itself.

The idea isn’t new. Knowledge-representation people have been working the shape for sixty years. What’s new is that language models can finally read it. Here’s the road — click through to see where the industry actually sits today.

step_01 LLM alone step_02 RAG step_03 GraphRAG step_04 Ontology RAG step_05 Tuned retrieval step_06 Self-describing step_07 Dynamic step_08 Closing the loop 2020 and earlier LLMs answer from their training data.

Fast, fluent, locked to what the model saw during training. No knowledge of your business, your customers, your deploy from Tuesday. Every wrong answer is stateless — the model can’t even tell you why.

Context Engineering has always felt like a better fit to me, but it just hasn’t seemed to gain traction, until perhaps now?

Daniel Davis · Context Graph Manifesto · Dec 2025 02 Anatomy

The unit: subject, predicate, object.

At the bottom of every context graph is the same three-part fact. Two things and the relationship between them. That’s it. Everything above this — ontologies, Cypher, Turtle, RDF, the entire stack — is what you do with many of these, connected.

Try it in context: Team Services Codebase Incident Learning Maya memberOf Platform

A person and the team they’re part of.

Subject Maya The thing the fact is about. Usually an entity — a person, a service, a commit, a concept. Predicate memberOf The relationship. Verb-like, specific. This is the piece engineers almost always underspecify — “is_related_to” is a tell. Object Platform Another entity, or a literal (a timestamp, number, identifier). Typed when it matters. 03 In practice

Where they’re already doing real work.

You’ve probably seen one of these this year without calling it a context graph. These are the shapes that show up most often when a team needs structure, not just search.

case_01 Agent memory case_02 Audit & provenance case_03 Multi-hop queries case_04 Temporal reasoning case_05 Ops & dependencies 04 Hands on

Build a context graph. Export it as YAML, JSON, Cypher, or Turtle.

Pick a domain. Add nodes. Connect them with predicates. The objective banner shows you what the graph is trying to capture at each step. When it’s shaped right, copy it out — that’s a working seed for a real context graph in TrustGraph, Neo4j, or any RDF store.

Graph builder Team & roles Service deps Incident trace Learning path Blank Represents People and the teams they’re part of. From side People, current roles. To side Teams, scopes, managers. From 0 nodes Add Before connecting: what does this relationship actually mean? That’s your predicate. Bad predicates are the #1 cause of bad graphs. Drag nodes to reposition Connect two nodes Connect To 0 nodes 0 triples Add Export YAML JSON Cypher Turtle Copy Reset

That export isn’t a toy. It’s the same shape you’d seed a TrustGraph context core with, or load into Neo4j with a LOAD CSV . The structure is the product.

You’ve already done the hard part 05 New craft

Three shifts that change how you ship software.

Context graphs don’t just unlock features. They shift what “good engineering” looks like. A few changes worth naming before the language settles.

shift_01

Structure is the product.

The ontology is no longer backend detail. It’s a design deliverable.

What entities does the system recognize? What predicates connect them? These decisions shape what users can ask, what the agent can answer, and what can be audited afterwards. A sloppy ontology produces a sloppy system no matter how good the model is.

shift_02

Reasoning leaves a trail.

Decisions become graph data. Every “why” is queryable.

When the agent’s chain of thought, the triples it looked at, and the model parameters in effect all live in the graph, you stop debugging with print statements and start asking the graph itself: why did we recommend this? This is the piece regulated shops have been waiting for.

shift_03

Context becomes portable.

“What the system knows” ships between teams like code.

Context cores — versioned bundles of knowledge, embeddings, policies — start to look like libraries. You build one, test it, tag it, promote it across environments. Institutional knowledge stops living in heads and starts living in packages.

06 Get started

A short path in, if this is new.

Two open-source projects worth your afternoon, and a five-item checklist to walk the first mile. Nothing below requires a platform contract.

Scaffold

create-context-graph

Neo4j Labs’ CLI. Pick a domain, pick an agent framework (PydanticAI, Claude Agent SDK, LangGraph…), get a running app with streaming chat, graph viz, and decision tracing. Four commands.

create-context-graph.dev → Platform

TrustGraph

Apache-licensed. Graph-native storage (Cassandra default, Neo4j supported), semantic retrieval pipelines, portable context cores you can version and promote like code.

trustgraph.ai → Reading

Context Graph Manifesto

Daniel Davis, 16 minutes. The clearest piece of writing in the space. Walks the full evolution, from someone who has been at the problem for eighteen months.

Read the manifesto →

The first mile

Name one dataset at work that is really a graph You already have one. Services and their deps. Tickets and the decisions that closed them. Commits and the authors + reviewers. The question isn’t whether the graph exists — it’s whether anything is using it as one. Write five triples about it by hand Subject, predicate, object. The point isn’t the triples — it’s finding out which predicates you don’t have good names for. That’s where the ontology work actually lives. Read the Context Graph Manifesto Sixteen minutes, no hype. You’ll finish with a clearer picture of where RAG stops and where context engineering starts. Scaffold a throwaway app with create-context-graph Pick a domain you know. Let demo data seed the graph. Poke at the decision traces. Four commands, nothing to uninstall cleanly. Decide whether you want structure or similarity Usually both. If you need both, a context graph stops looking like an option and starts looking like the default.

Context is the unit of work now.

If you’re thinking about where this fits for your team — what it replaces, what it doesn’t, what to try first — we’d be glad to talk it through. En Dash works at the seam between new technology and how organizations actually ship.

Start a conversation Try create-context-graph En Dash Consulting · Context Engineering · 2026