Your AI agent is rediscovering 85% of its context every run. Here’s the architecture fix (+ Contract Spec, Failure Triage, and Stack ADR)

Source: Nate’s Newsletter Date: 2026-05-13 URL: https://natesnewsletter.substack.com/p/rag-agents-knowledge-layer-architecture

Summary

Nate argues that production AI agents waste the majority of their compute re-assembling context on every run because they rely on vector search alone — treating retrieval as the whole solution rather than one component of a knowledge layer. The real architecture needed is a multi-component system combining retrieval, document structure intelligence, semantic data models, access control, provenance tracking, memory management, and write-back capability. Without pre-assembled context, agents make confident-seeming decisions from stale or incomplete information, producing wrong refunds, outdated policy citations, and corrupted metrics in production.

Implications

• Agent layer → lifecycle → orchestration thread. This piece diagnoses a structural failure mode in the current generation of production agents: the session and persistence layers are built, but the knowledge layer underneath them is missing. Dreaming, Auto Memory, and /goal all assume the context is correctly assembled — they’re orchestration sitting on a brittle foundation.
• Enterprise deployment as battleground thread. Nate’s frame directly maps to enterprise deployment failures. The 95% pilot-to-production gap (see the May 14 signal) and the SAP Dremio acquisition both become more legible: enterprises are buying data-layer infrastructure because agentic systems can’t function without pre-assembled, permissioned, provenance-tracked context.
• Context portability — “Memory is the moat” thread. The knowledge layer architecture is a technical specification for what the memory moat actually requires. Vendors who own the knowledge layer (pre-assembled context + write-back) own the switching cost, not just vendors who hold conversation history.