How Context Graphs and Decision Traces Are Transforming Enterprise AI
Introduction
In December 2025, Foundation Capital, a prominent Silicon Valley venture capital firm, released a paper titled “AI’s trillion-dollar opportunity” that has stirred considerable discussion in the enterprise AI community. The paper introduces two novel concepts: the context graph and decision traces. These ideas promise to address one of the most persistent challenges in AI—capturing the full reasoning behind business decisions. This article explores these concepts, their potential, and the broader implications for enterprise AI.
The context graph is essentially a specialized knowledge graph designed to store not just facts, but the entire context—including reasoning, causal relationships, and historical decision-making processes. According to the paper, “Agents don’t simply need rules; they need access to the decision traces that show how rules were applied in the past, where exceptions were granted, how conflicts were resolved, who approved what, and which precedents actually govern reality.” This insight has been echoed by industry commentators who note that the most valuable organizational knowledge lies in the data surrounding transactions and workflows.
What Are Context Graphs and Decision Traces?
At its core, a context graph is an evolution of the traditional knowledge graph. While standard knowledge graphs map entities and relationships, a context graph goes further by incorporating the how and why behind those relationships. Decision traces are the recorded paths of reasoning that led to specific outcomes—showing who made the decision, what alternatives were considered, and which precedents were applied.
This approach is powerful because it preserves institutional memory. In many organizations, critical decisions are made in meetings or emails, and the rationale is lost over time. Decision traces capture that rationale, making it accessible for future AI agents or human analysts. For example, a loan approval system could use decision traces to understand not just whether a loan was approved, but why—taking into account exceptions, risk assessments, and policy interpretations.
Jump to: Why All Three Are Essential
The Three Layers of Reasoning
The paper’s authors rightly point out that decision traces alone are not enough. To build truly reliable AI systems, we need to recognize that human reasoning relies on three distinct types of memory:
- Episodic memory – records of past events and decisions, i.e., decision traces themselves.
- Semantic memory – stored facts, concepts, and their relationships—the static knowledge of the world.
- Procedural memory – the skills and operational principles that guide how tasks are performed.
For enterprise AI, this means a context graph must integrate all three. Decision traces fall into the episodic category, but without semantic knowledge (e.g., what a “credit score” means) and procedural knowledge (e.g., how to calculate risk), the traces are incomplete. The Foundation Capital paper acknowledges this implicitly, but it’s worth emphasizing: a pure focus on traces risks missing the broader picture.
Why All Three Are Essential
Consider a scenario where an AI agent must decide whether to approve a contract. If it only has decision traces from past approvals, it might replicate past decisions without understanding why certain terms were acceptable. It might miss updated regulations (semantic gaps) or fail to apply the correct negotiation tactics (procedural gaps). The result: the AI may hallucinate—producing plausible but incorrect conclusions.
Serious AI requires integrating all three knowledge types. Skipping any one gives the system freedom to invent its own rules, which is a recipe for errors. As the paper suggests, context graphs are valuable, but they must be comprehensive. They should store entities, relationships, provenance (where knowledge came from), timestamps, permissions, policies, and—yes—decision traces. But not exclusively those traces.
To put it succinctly:
- Know the facts (semantic) – avoid reasoning without a factual foundation.
- Know the process (procedural) – understand how work is done, not just outcomes.
- Know the history (episodic) – learn from past decisions and their contexts.
This three-legged stool is the basis for reliable enterprise AI. Without it, organizations risk deploying systems that make decisions without accountability or transparency.
Conclusion
The Foundation Capital paper has rightly generated excitement by highlighting the importance of decision traces and context graphs. However, it is crucial to view these not as a silver bullet, but as part of a larger framework. Enterprise AI will succeed only when context graphs evolve into holistic knowledge stores that capture episodic, semantic, and procedural memory.
As the industry moves forward, companies should invest in building context graphs that incorporate all three reasoning layers. The promise is immense: AI that truly understands the logic behind business decisions, reduces hallucinations, and ultimately unlocks the trillion-dollar opportunity that Foundation Capital envisions. But that future depends on a complete, integrated approach—not a single magic key.