This might be the thousandth article written about context graphs in the last few weeks.
Enterprises are not ready for them.
Hear me out.
I recently read an article by Jaya Gupta, discussing context graphs as the next step for enterprise agents. The idea of a “context graph” is to collect decision traces over time, not just pieces of data. A trace records what choice was made, which rules or policies were applied and what outcome followed. When these traces are linked in a structured way, the context graph maps how decisions actually unfolded.
Jaya’s argument is thoughtful and I do agree with much of it. But I think we’re jumping several steps ahead of where most enterprises actually are.
The appeal of context graphs
Context graphs sound attractive for a simple reason. They promise continuity. Knowledge does not disappear. Decisions are remembered. Relationships between people, systems and events can be reused over time.
In theory, this is how agents become useful beyond single tasks. They start to act with awareness, not just instructions.
The problem is not the idea. The problem is where enterprises are today.
Enterprises are not structured for this yet
Context graphs rely on traces. Decisions, actions and outcomes need to be captured as they happen. In theory this sounds simple. In practice, those traces are hard to get and even harder to trust.
Most enterprises do not work like that. Information is scattered across tools. Decisions happen in meetings, chats and calls. Context lives in people’s heads and often leaves with them. Processes change faster than they can be documented.
Trying to build a rich memory layer on top of this reality does not create clarity. It captures noise.
Agents are already hard enough
Even without context graphs, deploying agents in an enterprise is difficult. Security rules, compliance checks, access controls, audits and internal trust all matter. These constraints determine what can be built and how fast it can move.
A context graph adds another layer because it introduces new data to capture, new links to maintain and new assumptions about what that data means. All of this has to sit on top of systems that are already complex.
The sequencing problem
There is a more basic issue here. We are trying to design memory for systems that are not yet producing consistently good work.
If the outputs today are messy, saving more of them does not help. It just stores confusion. Memory only adds value when there is something worth remembering.
Before we worry about long-term context, we should worry about the quality of daily work.
Start with people, not graphs
The first real step for enterprises is simpler: focus on helping people do better work using AI in their everyday tasks.
People should be able to improve the quality of what they produce. Clearer writing. Better organised thinking. Stronger analysis. Fewer avoidable mistakes. Work that stands up when reviewed.
They should also be able to get more done with less friction. Less time searching for information. Less repetition. Less manual rework. Faster completion of routine tasks. More time spent on judgement and decision making.
This does not require deep memory or rich organisational context. It requires tools that sit close to the individual and support the task at hand. Writing, rewriting, checking, summarising, comparing and explaining. Working with documents, spreadsheets, tickets, emails and code that already exist.
Productivity as a deliberate investment
Productivity should be a clear goal, not a side effect of experimentation.
Each role should be able to answer a simple question: how do I do better work, or get more done, with AI support?
When organisations invest here, they create better artefacts. Clearer decisions. More consistent outputs.
When context starts to matter
Once this foundation is in place, context starts to emerge naturally. There is something worth keeping. There are decisions worth remembering. There is structure that can be trusted.
That is when richer memory and more advanced agent behaviour start to make sense.
Context graphs may matter later.
For most enterprises today, the bigger gain is not smarter systems.
It is helping people do better work now.