Skill as Memory, Not Document: A Database-Native Substrate for Agent Skill Catalogs

Current LLM agent skill systems use storage formats optimized for human editorial workflow: Anthropic Agent Skills stores skills as SKILL.md files with YAML frontmatter and Markdown body; Voyager persists agent-authored skills as Python files. These document-first conventions collapse human authoring format, retrieval metadata, and agent-consumed runtime body into a single artifact. As autonomous agent learning becomes a primary use case — skills written by agents at runtime, retrieved by agents at inference, scoring composed over agent-emitted outcome events — document-shape choices become operational baggage.

I frame the required shift as "skill as memory, not document" and characterize three recurring failure modes when document-first catalogs are used as agent memory substrates: (i) token burn from in-context structural metadata; (ii) slowdown from filesystem organization not optimized for retrieval; (iii) invalid-skill admission from YAML-parseable acceptance. I measure all three on a 5,000-skill corpus.

Results: (i) Full-catalog disclosure at 5,000 skills consumes 919,200 tokens, exceeding GPT-4 Turbo 128K and Claude 3.7 200K windows entirely; the substrate's top-K disclosure consumes 369 tokens, constant in catalog size. A controlled ablation attributes the per-query gap against optimized filesystem alternatives to ~36 tokens of YAML frontmatter overhead per retrieved skill. (ii) p50 retrieval latency 87.3 ms, p95 106.3 ms at 5,000-skill scale on single-node deployment. (iii) The substrate rejects 70 of 70 adversarially-malformed skills (0% admission rate) at write time; a document-only YAML-parseable admission policy baseline admits 68 of 70 (97% admission).

Contributions: the "skill as memory, not document" framing as a unifying lens for autonomous-learning substrate design; the three-failure-mode taxonomy and its empirical measurement at 5,000-skill scale on a public reproducible corpus; and an operational decomposition of substrate-level primitives a competent filesystem alternative would need to assemble. I do not claim YantrikDB is a new database architecture — all individual primitives it ships (typed records, vector index, append-only logs, schema validation, Raft replication) exist independently in prior work. I identify a category mismatch in document-first skill catalogs and demonstrate a database-native design pattern for agent-written skill memory.