MCP live lookups won't scale; your AI needs a context graph

The core claim. Merge co-founder and CTO Gil Feig argues that companies need a "context graph" as the connective tissue for AI agents. He defines it as more than tool access: it combines third-party data, synced or cached data, memories, and skills that encode how a company actually works. The framing is useful even though Merge sells the infrastructure to build it.

The MCP limit. Feig's strongest point is that live lookups through MCP servers fail on aggregate or semantic questions. An agent can answer "Why was customer A upset last week?" by hitting Zendesk, but "Which customers were upset last week?" requires pulling thousands of tickets one page at a time until it times out. His conclusion: tools are not context; for analysis beyond simple lookups, teams need to sync a local copy into a vector database.

The layered design. He proposes four tiers: prompt, skills, and memory; live API calls; cached synced data; and derived context, such as stored summaries of past queries. A router sends a prompt to the right agent, the agent checks skills and memory first, then uses cached data if fresh within SLA, otherwise calls the live API, summarizes, and stores what it learned. The arrow is useful: encode company judgment as skills, not just data.

The skills example. Feig uses "which customer was upset" to show why context is not neutral. One company defines it as open tickets; another looks at Qualtrics survey data first because that is how the business works. That means teams must codify these processes as skills the agent checks before retrieving data. It is the most interesting idea and the least developed in the talk.

The open issues. Traceability gets real attention. Every answer needs source, fetch time, identity, permissions, transformations, and an invalidation rule. Feig says derived context should be stored so repeated questions do not re-query everything, but he does not explain how to keep derived summaries trustworthy. The examples are plausible, not production evidence. The talk also carries a clear vendor pitch: Gateway, Unified, and Agent Handler are Merge products.

The takeaway. The talk is more useful for its architecture than its novelty. Teams wiring one or two MCP servers will eventually hit the query scale problem he describes. The right question is not whether to build a context graph but how much of it deserves custom engineering versus buying a vendor layer. That decision is still open.