A new term is settling in the analyst vocabulary, and buyers will start arriving with it. In February 2026 Gartner described context graphs as the evolution of the knowledge graph for the age of agents: not a map of what an organization knows, but a graph of how it works — the decisions, the dependencies between teams, the informal paths work actually takes — kept so that an AI system can read it before it acts. The same research predicts that by 2028 a majority of enterprise AI systems will rely on one.
The word is worth understanding, because it names something real. It is also worth being precise about, because it will be stretched.
What the term means
A knowledge graph answers *what is true*: this product, this customer, this component, and how they relate. A context graph answers *why things are the way they are*: which decision led here, what it depended on, what it affected downstream. The defining property, in the analysts' own account, is the preservation of causal links — not only what happened, but why a decision was made and what it changed.
Two consequences follow. First, a context graph is built from the traces of work, not from documents about work. Tickets, commits, reviews, calendars, the routing of a request through people the org chart does not name: the graph is what those traces add up to. Second, it has a present tense. A document describes an intention at the moment it was written; a context graph is only useful if it still describes how the organization runs now.
Why it is arriving
Agents are the reason. An agent that drafts, summarizes or triages needs the same thing a new colleague needs on their first day: not the policy manual, but an understanding of how this place actually works — who decides, what depends on what, where the constraint sits this quarter. Without it, agents are fluent and wrong. With it, the same models become useful, and the difference is not in the model.
This is also why the term arrived from the human-work side rather than the data side. The context that matters most is organizational: decision patterns, dependencies across teams, the memory an institution keeps in its people. It is distinct from data context — what an agent can find — and from action context — what it may safely do. The organizational layer is the hardest of the three to obtain, because nobody writes it down.
What a memory adds
Padanet says *living organizational memory* in its own voice, and *context graph* where an analyst or an architect expects it. The mapping is honest, and it is incomplete in one direction. A living memory is a context graph plus two things the graph alone does not carry.
Evidence. Every claim in the memory opens onto the signals behind it, with its confidence and its recency attached. A graph can assert that a team sits on the critical path of work it is not resourced for; a memory shows the observations that assertion rests on, and how old they are. A claim without its evidence is an opinion with a good data model.
Purpose governance. A graph has an access model; a memory has a purpose model. Every read declares what it is for, and receives the smallest adequate, authorized, current set of claims for that purpose — what we call a context packet. An agent triaging tickets and a leader planning a program read the same memory and receive different things, because they declared different purposes. This is the part of the architecture that lets an organization run agents on its own data without turning work into surveillance, and it is the part a security review asks about first.
What the term does not mean
A context graph is not a graph database, though it may sit on one, and a buyer who hears the word "graph" should not conclude that they are being sold storage. It is not process mining, which reconstructs a documented process from logs and stops there. It is not an agent registry, and it is not a skills tool. It is the organizational layer of context — the answer to *how does this company actually work* — held current, with its evidence, for people and agents alike.
That is the thing worth building. The name will settle; the need already has.