In Brief
- Core Answer: A context graph is a connected record of the knowledge, state, history and decisions an AI agent needs to act well — entities and relationships, like a knowledge graph, plus the decision traces that explain why things are the way they are, each stamped with source, time and confidence. Where a knowledge graph answers "what is true," a context graph answers "what happened, why, and what should happen next."
- Why It Matters: Foundation Capital called context graphs "AI's trillion-dollar opportunity" in late 2025 on the argument that enterprise agents fail not from weak models but from missing context — the precedent, the exception, the reason. Neo4j, Glean, IBM and a wave of startups built the concept into products through 2026. For revenue teams, it is the layer that turns a CRM full of records into something an agent can reason over.
- Best For: Leaders and builders deciding how to give AI agents business context, anyone confused about context graph vs knowledge graph vs RAG, and revenue teams who want their agents to act on precedent rather than guess.
A context graph is a graph-structured store of everything an AI agent needs to complete a task in a business: the entities involved (accounts, people, products, deals), the relationships between them, the current state of each, the history of what has happened, and — the part that distinguishes it — the decision records that explain why. Neo4j's definition captures it: a connected representation of the knowledge, state, history and decision records an agent needs to complete a task. The idea gained its name and its momentum from Foundation Capital's December 2025 essay arguing that the durable layer in enterprise AI is not the model or the application but the graph of context that accumulates when every agent run leaves a trace — and that whoever owns that graph owns the compounding asset.
What a Context Graph Is
Three layers, stacked:
- Entities and relationships. The knowledge-graph layer. Account A is owned by rep B; contact C is the economic buyer; product D was sold in deal E. This is what a conventional enterprise knowledge graph holds.
- State and history. What is true now, and what was true before. Contact C was the champion until March; deal E moved from stage 3 to stage 4 on a date; the pricing changed in Q2. Every fact carries a validity period.
- Decision traces. Why. The exception that was granted and the reason; the discount that was approved and by whom; the account that was deprioritised because a competitor was entrenched. These are the records agents need to act on precedent instead of first principles, and the layer almost no company keeps.
Each statement in the graph carries metadata — source, time, confidence, access scope — which is what makes it trustworthy enough for an agent to act on. Graphwise's framing is useful here: a context graph is a knowledge graph in which every statement knows where it came from, how long it holds, and who may see it.
Context Graph vs Knowledge Graph vs RAG
The three are often conflated. They are different tools for different questions.
Approach | What it holds | Question it answers | Where it falls short for agents |
|---|---|---|---|
RAG (vector retrieval) | Chunks of documents, retrieved by similarity | "What do our documents say about X?" | No relationships, no time, no reasons; retrieves text, not facts |
Knowledge graph | Entities, relationships, attributes | "What is true, and how are things connected?" | Static; no history of change; no record of why |
Context graph | Entities and relationships, plus state over time, plus decision traces with provenance | "What happened, why, what is the precedent, and what should happen next?" | Requires capturing decisions — usually from agents — not just importing data |
Put the other way round — knowledge graph vs context graph — the knowledge graph is the static subset. It is the foundation of a context graph, and the site's knowledge graphs for GTM and sales AI guide covers why a graph beats flat CRM tables for relationship questions. The context graph adds the two dimensions agents most need: time, and reasons. RAG remains useful for retrieving supporting text but cannot answer a relational or a temporal question at all.
Context graph vs context engineering
Context engineering is the practice of deciding what goes into a model's context window for a given task — which facts, which history, which instructions. A context graph is the durable store that practice draws from. Without the graph, context engineering is rebuilt by hand for every agent; with it, the relevant subgraph is retrieved.
Why Context Graphs Became a Priority
The argument that made the term stick is about agent failure. Enterprise AI agents in 2025 mostly failed on context, not capability: they did not know that the account had an unusual contract, that this exception had been granted before, that the process has an extra step for a reason. Humans carry that context in their heads; agents cannot. The systems of record hold the "what" but not the "why," and the "why" — captured in Slack threads, emails and meetings — is never structured.
Foundation Capital's essay drew the conclusion: when agents run, they can leave a trace on every action — inputs considered, options weighed, decision made, outcome — and those traces, connected to the entities they concern, become the context graph. It is built as a by-product of work, not as a documentation project, which is why it can accumulate at a scale wikis never did. Dharmesh Shah's widely shared take was that the graph captures the "why" in the age of AI. Neo4j reframed it as three types of agent memory — semantic, episodic and procedural — living in one graph, and shipped it as a product story in 2026; Glean published a build guide; IBM added a definition. The concept moved from essay to category in under a year.
A Context Graph for a Revenue Team
Revenue is the natural first home for a context graph because the questions are relational, temporal and precedent-driven — and because agents are already running the process. Consider what a revenue context graph holds:
- Entities: accounts, contacts, buying committees, competitors, products, deals, campaigns, content assets, AI answer engines.
- Relationships: who influences whom, which competitor is in which deal, which asset was built for which opportunity, which answer engine cites which page.
- State over time: stage history, champion changes, pricing changes, share of recommendation in ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews week by week.
- Decision traces: why a discount was approved, why an account was deprioritised, why a piece of content was commissioned and what revenue it later touched.
With that structure, a deal-risk agent can see that the champion left and a competitor is multi-threaded before the rep does; a forecasting agent can weight deals by precedent rather than by stage — the case made in graph-based sales forecasting; and an attribution agent can trace revenue back through the asset, the opportunity and the decision that created it. That last link is what MultiplierAI's three agents write every cycle: Recon records what the answer engines say, Strategist records why each move was chosen, Closer records which asset served which opportunity — one database that is, in effect, a revenue context graph, and the reason the system learns per account instead of per prompt.
Context Graphs, Company Brains and Agent Memory
Three overlapping terms describe the same direction from different angles. A company brain is the business-level idea: one shared store of how the company works that people and agents both use. Agentic memory is the agent-level idea: the mechanisms by which an agent stores, organises and retrieves what it learns. A context graph is the data structure that serves both — the form the company brain takes when it needs to hold relationships, time and reasons, and the form agent memory takes when it needs to be shared. If the terms feel interchangeable, that is because in a well-built system they point at one thing.
The older discipline they all descend from is institutional memory — the knowledge that used to live only in people. A context graph is the first mechanism that captures it structurally, at the moment a decision is made, without asking anyone to write documentation.
How to Build a Context Graph
- Start from the questions agents will ask. "Has this exception been granted before?" "Who else at this account is engaged?" "What happened last time we discounted here?" The questions define the entities, relationships and traces you need.
- Import the entities you already have. CRM, product catalogue, org data. This is the knowledge-graph layer, and layering a graph over CRM records is a well-trodden path.
- Add time. Every fact gets a valid-from and, when it changes, a valid-to. Never overwrite; invalidate.
- Capture decisions from the agents. Instrument each agent to write a trace — inputs, options, decision, reason, outcome — linked to the entities involved. This is the layer that cannot be imported; it has to be generated by work.
- Attach provenance and access. Source, confidence and who may see each statement. Without these, agents will act on facts they should not trust or should not have.
- Govern writes. Decide what each agent may record, and review changes to the facts other agents act on.
Neo4j, Graphwise and the graph databases are the natural substrate; Glean's approach builds the graph over an enterprise search index; several startups sell the context graph as the product. For the outward-facing half of the problem — how a business is represented as entities in the answer engines' own knowledge graphs — knowledge graph optimization for AI search and entity SEO cover the mirror image.
Frequently Asked Questions
What is a context graph in simple terms?
A connected record of everything an AI agent needs to know to act well in a business: what the entities are, how they relate, what has changed over time, and why past decisions were made — each fact tagged with where it came from and when it was true.
What is the difference between a context graph and a knowledge graph?
A knowledge graph holds entities and relationships — what is true. A context graph adds state over time and decision traces — what happened and why — with provenance on every statement. A context graph is built on a knowledge graph; the reverse is not true.
Is a context graph the same as agent memory?
It is one form of it. Agent memory covers the mechanisms an agent uses to store and retrieve what it learns; a context graph is the shared, structured store that memory can live in when it needs to hold relationships, time and reasons across many agents.
Who coined the term context graph?
The term existed in graph and knowledge-management circles earlier, but its current meaning — the layer of decision traces and context for AI agents — was popularised by Foundation Capital's December 2025 essay "AI's trillion-dollar opportunity: context graphs," and adopted by Neo4j, Glean, IBM and others through 2026.
Why do revenue teams need a context graph?
Because revenue questions are relational (who influences whom), temporal (what changed) and precedent-driven (what happened last time), and agents running revenue processes cannot answer any of those from flat CRM records. A context graph is what lets a deal-risk, forecasting or attribution agent act on the actual history of an account.
References
- https://foundationcapital.com/ideas/context-graphs-ais-trillion-dollar-opportunity
- https://neo4j.com/blog/agentic-ai/what-is-context-graph/
- https://neo4j.com/blog/agentic-ai/context-graph-ai-agent-memory/
- https://www.glean.com/blog/how-do-you-build-a-context-graph
- https://www.ibm.com/think/topics/context-graph
- https://graphwise.ai/fundamentals/what-is-a-context-graph/
- https://www.linkedin.com/pulse/context-graphs-capturing-why-age-ai-dharmesh-shah-oyyze