# WORK ORDER — Station 7: the machine (GLAMMBRAIN) (GLAMMBOX pack)

This is the short version of Station 7. The full, step-by-step installer is
its own pack: `glammbrain-agent-pack.md`. Read this one first for the plain
map; hand your AI the other one to actually run the install.

## The real package now exists
A packaged, real download now exists: `glammbrain-20260827-full-v4.tar.gz`,
served from this same site at `/downloads/glammbrain-20260827-full-v4.tar.gz`
(sha256 `c7fde89a22ba289891041c25718891921312f107d0b1a9e7184f6147299c2723`).
Verify the hash before you run anything. Once it is downloaded and extracted,
its own `brain-engine/AGENT-PACK.md` is the authoritative install recipe and
supersedes the inline recipe in this pack and in `glammbrain-agent-pack.md` —
hand that file to your agent instead.

## Goal
Understand, in one plain sentence each, what a real memory stack is made of
before installing anything — then know exactly where the full recipe lives.

## The parts, one sentence each
- **Qdrant** (vector store): turns your notes into numbers and finds the
  passages closest in meaning to your question.
- **Neo4j** (graph, optional): maps who and what is connected to what, for
  questions a meaning-search alone answers badly.
- **Ollama + embeddinggemma:300m**: the local, free model that turns text
  into those numbers, running entirely on your own machine.
- **BM25** (keyword index): finds the exact word — a name, a model number,
  a code — that meaning-search alone tends to blur.
- **A cross-encoder reranker** (BAAI/bge-reranker-v2-m3, CPU fallback
  cross-encoder/ms-marco-MiniLM-L-6-v2): re-reads the top candidates from
  meaning and exact-word search and puts the best one first.
- **A feed script**: the one command that takes your folder of files and
  loads all of the above — nothing is indexed until you run it.

## What your computer needs
- Linux with systemd, or WSL2 + Ubuntu (Windows), is the verified installer path.
- macOS is a **guided beta**: use the same v4 archive, but give
  `pack-glammbrain-macos.md` to the coding agent before the archive's
  `brain-engine/AGENT-PACK.md`. The Linux service/timer installer is not a
  native Mac installer; background jobs need launchd or manual scheduling.
- Docker, installed and running.
- Ollama, installed and running.
- 16 GB RAM recommended (8 GB is a tight floor). GPU is optional — the
  reranker and embedder both have a CPU path, just slower.
- Windows native (no WSL2) is experimental — untested edge cases are likely;
  WSL2/Ubuntu is the proven path.
- The Mac path has not yet passed a clean-Mac end-to-end test. Keep that beta
  label until a real Mac installation receipt proves it.

## Prerequisites
- The home folder from Station 3, with real files in it worth searching.
- Docker and Ollama installed (the full pack below walks both).

## Steps
1. Read this station's plain map above so you know what you are about to
   install and why each piece exists.
2. Download and hand `glammbrain-agent-pack.md` to your AI agent — it is the
   complete, runnable recipe: Qdrant plus optional Neo4j, the Ollama
   pull, the Python environment, the collection setup, the BM25 index, the
   reranker service, the feed script, and a verification search.
3. Let the agent run it inside your home folder, asking before it writes,
   exactly like every other station on this trail.
4. Confirm with the verification test in that pack: ask one question you
   already know the answer to, and check that the answer names the source
   file.

## Verification
- You can explain each of the six parts above in one sentence, unprompted.
- You have `glammbrain-agent-pack.md` saved and ready to hand to your agent.

## Honest costs
Every piece named here is free and open source. Disk space, RAM, and
patience are the only costs — this pack ships EMPTY, with none of your data
in it; you feed it yourself. The packaged tarball is MIT-licensed and free to
keep, fork, or hand to anyone; there is still no public GitHub repo — the
download above and this pack's companion are the full recipe today.