Where the idea comes from
A clear definition comes from Ashu Garg and Jaya Gupta of Foundation Capital, in the essay “AI’s trillion-dollar opportunity: Context graphs”. Their argument starts from a gap in today’s software. Systems of record like a CRM, ERP or HR system store the current state. In their example, the record says a customer has a 20% discount, without linking it to the approval and policy exception behind it. The reason may be missing from the record or scattered across notes and messages. The reason lives in a Slack thread, a call, an email to Finance, or someone’s head.
Their summary: “The wall isn’t missing data. It’s missing decision traces.”
The two building blocks
| Term | What it is | Example |
|---|---|---|
| Decision trace | A record of one concrete decision: inputs gathered across systems, the policy applied, the exception granted, who approved, what was written | A renewal agent proposes a 20% discount, routes the exception to Finance, and records the reason |
| Context graph | The structure that forms as traces pile up, linking accounts, tickets, policies and approvers through decision events and “why” links | Next time a similar account asks, the agent finds that precedent and the reasoning behind it |
The authors stress that the graph is “not ‘the model’s chain-of-thought,’ but a living record of decision traces stitched across entities and time.”
Why it matters for AI agents
An AI agent that only sees current state can tell you what the discount is. It cannot tell you whether the next customer should get one too. Rules describe general policy; traces show how that policy was applied in practice, including the exceptions that are the real policy in most companies.
This is context engineering at the scale of an organization. Anthropic’s advice for agents is to keep light references in context and retrieve detail just in time. A context graph is a good thing to retrieve from: instead of pasting a whole history into the window, the agent pulls the two or three precedents that match this case.
The common misunderstanding
The most common mistake is to think a context graph is a data project: connect all systems, pour everything into a graph, done. Foundation Capital’s own point argues against this. Warehouses receive data after the decision, through ETL, when the context is already gone. The traces have to be captured when the decision is made, in the workflow itself.
That makes recording, not storage, the hard part: capturing decisions accurately and consistently while the work happens. A graph full of half-recorded reasons is not much better than none.
A second misunderstanding is that this only matters for large enterprises. Any team that makes exceptions, corrects its AI, or has “the way we do it here” has decision traces. Most of them are lost in personal chats.
What it means for a team
In a team that uses AI, decision traces are created every day and thrown away every day. Someone corrects ChatGPT on how a client wants invoices; someone tells Claude Code why a module is built the odd way it is. That correction stays in one person’s session. The next colleague, or the same person in a different tool, starts from zero. A shared context graph can make reviewed corrections available as precedent to everyone’s AI, with their source, where they apply and who may see them.
For the broader picture, see what is context engineering and how to set up context engineering for a team.
Where Bombos fits
Bombos is a context graph for teams that already use AI: it links team knowledge, learned rules and what already lives in mail, documents and repos, and retrieves from it by similarity, keywords, nearness in the graph and rules that only apply in a given context. Every AI tool the team uses, from Claude and ChatGPT to Copilot, Cursor or your own agents, reads from the same graph, and you can trace what each answer rests on. It runs in the cloud or on your own server, and one person can start with their own agents.
Related questions
Is a context graph the same as a knowledge graph?
Not quite. A knowledge graph usually links facts about entities, such as this customer belongs to that region, though it can model decisions too. A context graph, as Foundation Capital defines it, centers on decision traces and precedent: why this customer got that exception, and who approved it.
Is a context graph the model's chain-of-thought?
No. Foundation Capital is explicit about this: it is a living record of decision traces across entities and time, not the reasoning text a model produces.
Where does the data for a context graph come from?
From the moment decisions are made: approvals, exceptions, mail threads, tickets, the reasons people give. That is why the authors argue it is best captured by whatever runs the workflow, not by a warehouse that receives data afterwards.
Do I need a graph database to build one?
Not necessarily. The idea is about what you record and how it links, not about a specific database. Many setups combine a graph with similarity and keyword search.
