What context graphs still get wrong about organizational knowledge
Context graphs capture decision traces. They do not tell you when an old decision stopped applying, whether two competing decisions are both live, or what remains unresolved.
Context graphs have become a serious concept in enterprise AI infrastructure. The basic argument is right: a flat retrieval index does not capture how decisions relate to each other, how policies evolve, or which precedents are still active. A graph that models decisions as first-class nodes, with edges representing temporal validity, dependencies and supersession, gives AI agents a richer substrate to reason from.
Foundation Capital put forward the investment thesis in a December 2025 essay by Jaya Gupta and Ashu Garg, and Thoughtworks added context graphs to their Technology Radar in April 2026. That level of attention is not wrong. Decision traces are a real gap that retrieval does not fill.
But context graphs have a specific blind spot, and it matters for any organization trying to build AI agents that give reliable answers over time.
What context graphs do well
A context graph captures that a decision was made: who made it, when, what it applied to, what evidence it was based on. It can model the relationship between a policy and the exceptions to it, between a current ruling and the precedents it overrides.
This is genuinely better than treating the same information as unstructured text. An agent reading from a context graph can see that a pricing exception was granted, who approved it, and when it applies. It can follow a chain of policy updates to find the current state.
The critics who say “retrieval is not enough” are right. Text ranked by similarity cannot represent these relationships. Graph structure can.
The question context graphs do not answer cleanly
Here is the specific problem.
A pricing exception was granted under the previous CFO. A new policy was introduced three months later. Does the old exception still apply?
The context graph has both nodes. It has the relationship between them. What it often lacks is a formal representation of whether the old node is still active. The graph stores that both exist. It may store that the new policy was introduced after the exception. What it does not store, unless explicitly designed to, is: the old exception was superseded by the new policy, and here is the decision that made it so, with its scope and rationale.
This is the supersession gap. Supersession is not just “a newer thing exists.” It is the specific act of the organization deciding that an older conclusion no longer applies, with an attributed, evidence-backed record of that decision and its scope.
Without a supersession mechanism, an agent reading from a context graph must infer whether the old decision still applies. Inference is exactly the wrong tool for this. The organization made a decision. That decision should be represented as a decision, not something to reason about from surrounding context.
Conflict is a different problem
Related but distinct: what happens when two live decisions are genuinely incompatible?
Finance counts a metric one way. Product counts it another. Both are current. Neither has been superseded. The organization has not yet decided which definition to use in the new context.
A retrieval system averages or picks one. A context graph can represent both decisions as nodes, but it does not know they conflict unless the conflict is modeled explicitly. Even when it is, the graph structure does not tell the agent what to do: give the finance answer? the product answer? an average? or surface the conflict and decline to answer?
The right answer in most cases is to surface the conflict. An agent that confidently gives one number when the organization has not yet resolved a conflict is providing false certainty. The conflict is part of the organizational state. It should be visible.
Conflict as first-class state is different from conflict as a graph query. A graph query can find two inconsistent nodes. But representing conflict as a standing state, one that the organization formally acknowledges as open and routes to a reviewer, requires a different model. The conflict is not just a logical observation. It is an organizational fact that someone needs to act on.
The unresolved state
Beyond supersession and conflict, there is a third case: the organization has not yet made a decision. A question was raised. The relevant people disagree. The answer is genuinely open.
This is different from a conflict between two live decisions. This is a recognized gap. The organization knows it needs to decide and has not done so yet.
An agent asked a question that depends on an unresolved organizational issue should not give a confident answer. The honest answer is: this is not settled. Here is what is known, here are the competing views, here is why it has not been resolved. That is useful information.
A context graph can store everything that has been said on the topic. It cannot easily represent the absence of a decision as a standing state that the agent should treat as a signal. Unresolved is not “no relevant nodes.” It is a positive organizational state: we know we do not know, and someone is responsible for getting us to a decision.
What a maintained organizational understanding needs to represent
Pulling these together, a system that gives AI agents reliable answers about organizational knowledge needs to represent four distinct states for any given conclusion:
Current. This is what the organization formally holds right now, grounded in evidence, approved by the right person.
Superseded. This was the held position until a specific point. Preserved with the evidence, the decision and the scope of the change. Still queryable for historical questions.
In conflict. Two or more positions are live and incompatible. The conflict is a standing state. Answering questions that depend on it requires surfacing the conflict, not picking a side.
Unresolved. A decision needs to be made. The organizational process has not yet produced one. Questions depending on this should return the fact of unresolution, not an inference.
These four states are structural. They are not annotations an agent reasons about. They are properties of the organizational record that determine how a question should be answered.
Why this is hard to retrofit onto a context graph
A context graph stores what happened. Adding supersession means adding the act of supersession as a modeled event, with its scope and rationale and authorization. Adding conflict as a standing state means adding a process for recognizing conflicts and routing them to resolution. Adding unresolved as a standing state means adding a process for recognizing open questions and tracking them.
None of this is a graph schema problem. It is an organizational governance problem. The graph can store the result of a supersession decision. What it cannot do, by itself, is ensure that supersession decisions get made and recorded, or that conflicts get recognized as conflicts rather than treated as inconsistent data to query around.
This is why the infrastructure question is not just “how do we represent knowledge as a graph?” It is “how do we maintain the standing of organizational conclusions over time, so that agents draw on a current state the organization has actually approved?”
Where the argument for context graphs understates the challenge
The strongest versions of the context graph argument recognize this. Foundation Capital’s essay notes the open question of whether capturing true intent is feasible. The InfoWorld analysis asks whether decision traces are enough. These are the right questions.
The answer is that decision traces are necessary but not sufficient. Capturing the decisions made is the first step. Maintaining their standing, routing conflicts to resolution, representing open questions as open rather than as data to interpolate, and making supersession a formal act rather than an inference: these are the steps that turn a record of decisions into a governed current state.
For organizational knowledge to be reliable for AI agents over time, the infrastructure needs to represent not just what was decided, but what is currently authoritative, what is no longer authoritative and why, what is contested, and what has not been decided yet. These four states are the difference between a decision archive and a maintained understanding.
Propperly maintains the organization’s analytical conclusions as a governed current state with these four properties. See what that means in practice.
See how Propperly maintains organizational conclusions in four states.
Start with a bounded reconstruction of one domain using your own definitions, decisions, and dashboards.