Cal Paterson proposes storing agent memory as a zip of Markdown pages

Cal Paterson published a proposal for agent memory that fits in a paragraph. A memoryfield is a zip file:

my-memories.memoryfield.zip
├── carbon-fibre-woks.md
├── finnish-bureaucracy-tips.md
├── [... many more md files ...]
└── nomic-embed-text-v1.5.sqlite3

Markdown pages, optional YAML frontmatter, and an optional SQLite vector index for semantic search. That is the whole format.

The argument against the alternatives

He groups current memory systems into three kinds, and the criticism of each is specific.

The first ties you to one harness, usually written by the lab selling you access to that harness, and works by reading through your conversation history. The result is that most stored memories are about you, when information about the world is generally more useful.

The second is elaborate. He describes one prominent system needing pgvector, a Neo4j graph database and a language model of its own purely to decide what is worth remembering. Beyond the administration cost, his claim is that the complexity confuses the models and does not scale as the models improve.

The third imagines an idealised rational memory, which means a graph and sometimes logical propositions. That strips information out of its context and leaves it isolated. A list of short extracted facts is not much use to anyone.

What all three share, in his framing, is treating memory as a process. His position is that memory is better represented as data, and he quotes Brooks on it: show me your tables and I will not usually need your flowcharts.

Two design decisions worth copying

Write prose, skip chunking. Retrieval pipelines get complicated because they are making existing human documents readable by a machine, and those documents are often large PDFs. A memory is different: at the moment it is created, the agent that has it can write perfectly good Markdown. No chunking, no enrichment, no double summarisation. Let it write the memory directly.

A size limit that is a feature. A page has to fit in a vector embedding, so there is a soft limit of about 8KB, roughly 2,000 tokens or 1,300 words. He argues that is a restriction worth having anyway. Need more detail, add more pages, which agents have no trouble doing.

A page carries title, created, updated, uuid and summary in frontmatter, then prose.

Why this is worth twenty minutes

Read it as a design argument rather than as a product, because it is one.

The properties you get automatically by choosing a file format are the point. A memoryfield can be opened in a text editor, committed to git, diffed in a pull request, copied to another machine, and read by a person who wants to know what their agent thinks it knows. None of those need a feature to be built.

Whether it works at scale is unproven, and the post does not claim otherwise. The question it puts to anyone currently setting up a vector database and a graph store for agent memory is a fair one: what does that give you that a folder of Markdown does not?

Source

• Agent memory as a file format, Cal Paterson, August 2026