- Author
- Cal Paterson
- Published
- August 2026
- Kind
- Personal blog post, with a spec and tools
- Original
- calpaterson.com/memoryfields.html
- Subject
- File-based agent memory
- Themes
- What lasts, Given away
What it is
Cal Paterson proposes memoryfield, an open file format for agent memory. Paterson provides a written spec, a command line tool and an agent skill. The post presents it as a simpler alternative to existing memory systems.
What it says
- Memory should be data rather than a process.
- A memoryfield holds Markdown pages, optional YAML frontmatter and an optional SQLite vector index.
- Pages have a soft limit of about 8 KB, roughly 2,000 tokens.
- Agents write prose directly, with no chunking, re-ranking or hybrid search.
- Semantic search jumps to relevant pages, so recall takes two tool calls at most.
- Paterson argues that walking linked wikis is slow and misses relevant pages.
- Memories work best with citations, ideally URLs.
What evidence it gives
The case rests on design reasoning. The post reports no benchmark results. Paterson mentions observed agent behaviour, such as using perl across a corpus. The two-call figure follows from the design and is not a measurement.
Interests and limits
Paterson wrote the format and its tools, so the case is not neutral. Paterson says irrelevant memories do no harm because search never surfaces them. The post gives no test of this. Paterson warns against sharing memories with untrusted parties. The post also notes that installing the tools takes four package managers.
Connections
These links are this wiki's own reading, not claims the sources make.
- What Heraldry and Japanese Mon Can Teach Us About Building Visual-Identity Generators. Overmyer lets a state machine run generation while a tree holds the design. Paterson keeps memory as data, apart from any pipeline. Both separate what lasts from the process that produced it.