BrookeOehm Smith

2026 — ongoing

Herd

A retrieval system for writing sessions that run for hours, so the conversation stops forgetting what the essay says.

Private repository — the writing here describes the approach rather than linking to code.

  • TypeScript
  • RAG
  • semantic retrieval
  • SQLite

Writing something long across many files with a chat assistant means re-uploading material every hour, splitting the work across threads, and losing the thread of the argument. The cause is simple: the context window fills, and everything earlier quietly stops counting.

Herd indexes an essay project once, retrieves what is relevant to each turn by semantic similarity, and holds the draft you are actually working on in full context rather than letting it be summarised away. Session history is managed by a rolling window with summarisation, so a conversation that runs for hours degrades gracefully instead of falling off a cliff.

The active-draft pinning is the part that makes it usable. Retrieval alone gives you an assistant that knows roughly what your essay contains; pinning gives you one that knows exactly what the paragraph in front of you says.

Sibling to Cattle — Herd is where an essay is written, Cattle is what happens to it afterwards. The repository is private.