From Retrieval to Evidence: A Task-Adaptive Two-Layer Knowledge Architecture for LLM Agents

Retrieval-augmented generation and structured agent memory are commonly treated as competing ways to ground large language models. Our results show that their relative value changes with the task and that neither retrieved similarity nor graph association is sufficient evidence for an answer. We present a task-adaptive two-layer methodology combining a compact, reviewed knowledge layer for orientation and cross-source synthesis with a versioned evidence layer for exact lookup and claim support. A router selects the representation, while an independent admission stage checks named entities, time, source role, contradiction, and provenance. A controlled course-grounding study reveals a task-conditional crossover within Course 2: the distilled Body of Knowledge (BoK) trails plain RAG by 1.07 points on lookup and leads by 0.62 points on synthesis. Across a 72-question three-course lookup panel, RAG leads by 2.17 points; Course 1 synthesis is near parity. Complementary, position-debiased LLM-judged mastery comparisons favor the BoK on both courses. The standing representation is reduced 22–35-fold. An observational production case over a 289,642-chunk evidence base illustrates why candidate discovery, packet selection, and evidence admission require separate gates. The contribution is a tested method for assigning retrieval, structure, and evidence admission distinct responsibilities within one grounded agent architecture.