Cal Paterson proposes agent memory as a file format of Markdown pages and a SQLite vector index

Sources: primary, discussion
Summary: Paterson's proposal treats agent memory as a file format rather than as a retrieval pipeline: Markdown pages on disk, with a SQLite vector index over them for lookup. The argument is that agents already handle Markdown, bash and SQLite directly, so a store defined as files needs no fixed API between the agent and the data.
Why it matters: The argument is that a data format beats a pipeline, because agents already handle Markdown, bash and SQLite, and a store with no fixed API does not need rewriting as models improve.