A consistent pattern across knowledge graph architectures: the moment a system grows beyond simple file storage, it discovers that narrative depth and structured querying pull in opposite directions. Pal built a wiki for depth alongside PostgreSQL for cross-dimensional queries. Hermes uses curated MEMORY.md alongside SQLite with FTS5 alongside HRR holographic vectors — three layers, each serving a different retrieval shape. Our own graph mirrors this with graph-index.yaml (structured metadata for routing) alongside full insight files (rich prose for understanding).
The question isn’t “which storage?” — it’s “how do you route between layers?” Pal solves this with prompt-engineered metadata routing. We solve it with a CLAUDE.md instruction that directs agents to consult graph-index.yaml first, then drill into full files only when needed. Both are imperfect: Pal’s routing depends on LLM judgment per query, ours depends on the agent remembering the instruction.
This connects to Structure plus reasoning beats flat similarity for complex domains — the dual-layer pattern IS the structural precondition for reasoning-based retrieval. You can’t reason about which storage to query if everything lives in one flat layer. It also explains why Files are the universal interface between humans and agents works well as one layer but needs complementary structure: files excel at narrative depth but can’t answer “everything about topic X from the last two weeks” without an index layer on top. The The three-layer AI stack: Memory, Search, Reasoning pattern describes the same tension at application scale — memory, search, and reasoning each need their own substrate.