KM4BI KM4BI

SAP bets over a billion on the agentic RAG problem many teams miss

Agent memory is moving beyond vector search: production agents need governed context, source authority, permissions, and task-shaped data bundles.

Production agents are exposing a limitation in classic RAG. The old chatbot pattern — retrieve a few semantically similar chunks and answer a question — does not fit agents that must act across records, policies, contracts, metrics, and permissions.

What is changing

Pinecone’s Nexus and NoQL are framed as a sign that vector similarity is not enough by itself. Agents need operational context: intent, filters, provenance, access policy, confidence, response shape, and budget. Page Index takes a different route by arguing that many documents should preserve their hierarchy because structure is part of meaning.

Why SAP matters here

SAP’s moves around Dremio and Prior Labs point to the enterprise version of the same problem. Critical business knowledge often lives in governed tables, ERP systems, CRM records, metric definitions, and semantic layers rather than in prose. For an enterprise agent, lineage and permissions are not extras; they determine whether the answer is usable for action.

The practical takeaway

The recommended order is: define the contract between the agent and the data, write down the exact bundle the agent needs, then choose the primitives that can deliver it. That may mean vectors for prose, document trees for structured documents, semantic layers and tabular reasoning for business data, and graphs for relational knowledge.

Bottom line

The memory race is not about adopting the trendiest retrieval tool. It is about delivering appropriate context in a form the agent can reliably use. Bigger context windows help less than disciplined data contracts.

Source