Cal Paterson, a software engineer, published a proposal on August 31 for storing AI agent memory as a plain file format rather than a database or pipeline. He calls it memoryfields: a zip archive of Markdown pages, each with optional YAML frontmatter, indexed by a SQLite file holding vector embeddings for semantic search.
Agent memory is currently the stickiest lock-in point in the AI stack, because it is the one asset a user actually accumulates and cannot take with them. Chat history, learned preferences, and project context pile up inside a single vendor’s harness, and switching labs or agent frameworks means starting over. A plain-files format is as much an argument about who owns that accumulated context as it is a technical proposal.
Paterson’s critique targets three existing approaches. Memory systems built by model labs, he argues, mine conversation history and mostly capture facts about the user rather than the world, serving the lab’s push from selling API access into selling a platform. A second category relies on heavy infrastructure, and he points at one unnamed system whose retention decisions require a Neo4j graph, pgvector and a language model of their own. A third category, which he labels “High Modernist,” forces memories into knowledge graphs or logical propositions, stripping the context that makes a memory legible.
His alternative treats memory as data, not a pipeline. A memoryfield holds Markdown pages capped at roughly 8kb, or about 1,300 words, a limit set by what fits comfortably into a single vector embedding. Paterson frames graph-based systems, which he traces to Andrej Karpathy’s concept of hyperlinked “wikis” for agents, as too slow: retrieving information N steps deep in a knowledge graph costs N+1 sequential tool calls, each adding two to three seconds of latency. Semantic search over the flat file set, by contrast, needs at most two calls: one to search, one to read the matched pages in parallel.
Paterson also argues that a low-mechanism system scales with model capability in a way a fixed API cannot. Agents already write competent Markdown and bash, he says, and he has seen them improvise access patterns unsupported by any interface, such as running Perl find-and-replace across a memory corpus or embedding CSV data that gets queried with SQLite. A heavily engineered memory API, by comparison, only supports the operations its designers anticipated.
The proposal ships with a published spec, a command-line tool, and an installable skill, but it remains one engineer’s design, tested primarily on his own agents and a demo archive called soapstones.memoryfield.zip. Paterson has not published adoption data or head-to-head benchmarks against the graph-based or database-backed systems he critiques. The embedding model he defaults to, nomic-embed-text-v1.5, is small enough to run without a GPU, which matters for anyone trying this locally rather than through a hosted vector database.
Teams building agent memory in house should treat memoryfields less as a standard to adopt and more as a test of a specific claim: that portability and inspectability beat retrieval sophistication once an agent’s memory grows past a few hundred entries. Anyone evaluating a vendor’s memory feature should ask whether the resulting data can leave in a form as simple as a folder of Markdown files, because that answer says more about the vendor’s incentives than any benchmark will.
Cal Paterson described the memoryfield format and its rationale in a post published August 31, 2026, on his website, calpaterson.com.
