← @bgSponge

Agent memory as a file format

Summary

Cal Paterson argues agent memory should be a portable file format, not a harness-tied pipeline.

  1. Memory is data, not a process.

    Lab-tied mining, graph-plus-vector stacks, and distilled-fact stores all put machinery between the agent and what it needs to remember.

    But memory - especially to a model - is much better represented as data.

    Memory is data! The less fixed machinery we put between the agent and that data, the better the agent can be.

  2. Write prose pages, not chunks.

    The agent authors Markdown memories itself, so they do not need RAG-style chunking; keep each page under about 8kb so it still embeds.

    just have the agent write the memory directly in its favourite format (which is Markdown).

  3. Jump by content, do not walk the graph.

    Semantic search plus parallel reads find relevant pages in two tool calls without serial graph traversal or SEO-style titles.

    So in a memoryfield, at most 2 tool calls are required (#1 to search, #2 to read in parallel).

  4. Low mechanism scales with the model.

    A zip of files plus a deletable SQLite index lets agents invent bash, perl, and SQL access patterns as models improve.

  5. Keep the format open and pinned.

    Portable zip archives, a spec, and sha256 review avoid harness lock-in and let untrusted shared memories be inspected.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →