Celent has published a new report, entitled Knowledge Management in Agentic Systems — Part 1: Foundations.
Part 1 of the four-part series lays the conceptual groundwork the rest of the series builds on. Specifically, it covers what “knowledge” actually means for an AI agent. And why treating it as one thing rather than four — procedural, semantic, episodic, organisational — is the most common design mistake that Celent notes in its research. It examines the working difference between retrieval-augmented generation and its graph-based variant, GraphRAG, and the layered memory architecture (working, episodic, semantic, procedural) that determines whether an agent’s behaviour holds up in production or degrades into inconsistency.
It closes with the strategic argument that gives the series its throughline: model capability is commoditising quickly, and the durable competitive advantage sits in the knowledge and context layer a firm builds around it — which is why the report series argues for model-agnostic architecture from the outset.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
Why this topic, what is Celent’s angle?
Beattie observes that every institution he speaks with — regardless of vertical — is now running some version of the same conversation. That is, which model to adopt, and how to evaluate the next one when it arrives six months later. Almost none of them are asking the more durable question, which is what they’re building underneath the model that will still be valuable when today’s model is obsolete. Beattie argues that it is that the knowledge layer, not the model, that is the asset worth architecting deliberately, and that most current agentic AI deployments are brittle precisely because they’ve skipped that step.
The strategic conclusion is a direct challenge to how most firms currently frame their AI investment. Models are commoditising fast. A well-governed knowledge layer compounds in value the longer it runs and the more it is used. The firms that recognise this now — and build accordingly — will own something in two or three years that a competitor cannot simply buy off the shelf.
Future instalments examine knowledge graphs and regulatory representation by industry vertical, the deployment evidence from named financial institutions, and the vendor landscape for agentic memory infrastructure.
The report will especially be of interest to CIOs, CTOs, enterprise architects, and heads of AI/data strategy — across wealth management, banking, insurance, and capital markets.
Further information on how to access the full report is available via this link.
