Cal Paterson proposes memoryfields, agent memory as a zip of markdown pages plus a SQLite vector index

Cal Paterson·medium signalXBlueskyLinkedInCopy link

Paterson argues agent memory should be a portable file format rather than a pipeline, and specifies one: a .memoryfield.zip of markdown pages with optional YAML frontmatter (title, created, updated, uuid, summary) and a SQLite embedding index for semantic jump. He caps pages at about 8KB so each fits in a single embedding, treating the limit as a feature that forces more pages instead of longer ones. His case against Karpathy-style knowledge-graph wikis is operational: retrieving information N hops deep costs N+1 serial tool calls at 2-3 seconds each, link text is a weak relevance signal that pushes authors into SEO-style title hacking, and walking the graph dumps irrelevant pages into context.