Engineer Cal Paterson has published a widely discussed essay arguing that agent memory should be a file format, not a multi-stage pipeline. Titled 'Agent memory as a file format', the piece takes aim at the three dominant approaches to AI agent memory: vendor harnesses that mine conversation history, overengineered stacks that pair vector databases with graph databases and a dedicated LLM, and high modernist graph-based designs. All of them, he argues, treat memory as a process, when memory is much better represented as data.
His proposal is a portable format he calls a memoryfield: a ZIP archive of Markdown pages with optional YAML frontmatter and an optional SQLite vector index for semantic search. Agents write memories directly as prose, which removes the need for chunking, enrichment and double-summarisation that makes RAG pipelines complicated. Pages are capped at around 8KB, roughly the length of a magazine article, and agents simply add more pages for more detail.
The essay also argues against knowledge-graph traversal for retrieval, pointing out that walking a graph costs N+1 tool calls for N-deep paths and pressures agents into SEO-style link text. The post has sparked a lively debate on Hacker News, with commenters weighing in on what production agent memory should actually look like.