**Summary** Memoryfield is a file-format specification for agent memory: a zipped collection of Markdown files with optional YAML frontmatter and an SQLite vector index for semantic search. The core claim is that memory should be *data*, not *process*—agents write prose directly to `.md` files (~8kb soft limit each), and retrieval happens via one semantic search call plus one parallel read, rather than via graph traversal or complex extraction pipelines. The author argues this beats existing systems because (1) prose preserves context better than "distilled facts"; (2) semantic search is faster and more reliable than link-walking; (3) agents can invent their own access patterns (bash, SQL, Perl) as they improve; (4) the open format and simple zipfile transport avoid vendor lock-in. A CLI tool, embedding model (nomic-embed-text-v1.5), and agent skill are provided. The practical claim: simpler, more flexible agent memory that scales with model capability. The source doesn't quantify performance gains, user adoption, or failure modes in real deployments—it's a specification and design argument, not an evaluation.