Multi-Layer Knowledge Atoms: An Agent-Native Knowledge Format with Human-Readable Fallback

LLM agents waste tokens by consuming full documents when curated summaries would suffice. We propose the Multi-Layer Knowledge Atom (MKA) — an agent-native knowledge format where YAML frontmatter is the primary interface for machines, and Markdown body with HTML comment markers (<!-- @scan/read/deep -->) provides human-readable fallback at three resolution levels. Benchmark (N=300, commercial LLM, cross-validated): scan-layer retrieval achieves 93% of full-text keyword recall at 76% fewer tokens (3.9x efficiency). A reasoning benchmark quantifies the gap beyond keywords: scan retains only 28.4% of reasoning-relevant information, with negation/caveat awareness dropping to 8.3% — validating the multi-resolution design. The reference implementation includes budget-aware retrieval, graph traversal over typed relations, context-aware conflict resolution, layer integrity checking, and reasoning benchmarks. 235 tests, 16 modules, 14 CLI commands, 6 MCP tools. Defensive publication establishing prior art for: (1) the <!-- @scan/read/deep --> multi-resolution marker system in Markdown, (2) budget-aware knowledge retrieval with progressive resolution upgrade, (3) graph traversal within token budgets over typed knowledge relations, and (4) context-aware conflict resolution distinguishing true from conditional contradictions.