ZERO TO A REAL AGENT. SEVEN STOPS.

Each stop below ends with a file you can download. Hand that file to your own AI — Claude, Gemini, or any agent installed on your machine — and it contains the procedure, step by step, for the agent to carry out.

Two doors are enough to start: ChatGPT (from OpenAI) or Claude (from Anthropic). Both have a real free tier and a real paid tier. Pick one — you can always try the other door later.

Sign up here — the official doors, nowhere else:

• ChatGPT: chatgpt.com — official pricing: openai.com/chatgpt/pricing
• Claude: claude.ai — official pricing: anthropic.com/pricing

We do not print a dollar figure here — prices change. Read today’s number off the official page yourself, every time.

Same account, two apps: the PC app for real work, the phone app so it is with you the rest of the day. Both stores, first login done by you.

• ChatGPT desktop: openai.com/chatgpt/download — phone: search “ChatGPT” by OpenAI on the App Store or Google Play.
• Claude desktop: claude.ai/download — phone: search “Claude by Anthropic” on the App Store or Google Play.

A plain chat window sees only what you type or upload into it — not your desktop, not your other apps, not your files. That changes the moment you give it a folder. Understanding that line is the whole point of this station.

One folder, on purpose — never the whole Documents drive, never Downloads, never the home directory “to be safe.” A scoped agent is a safer agent, not a slower one.

A chat answers. An agent inside an editor can work on files you deliberately open. Start with one small, reversible task; it may help with later setup only after you review its work and approve each permission.

• VS Code: code.visualstudio.com
• Claude Code documentation: docs.claude.com/en/docs/claude-code
• Codex second, if you signed up with ChatGPT: openai.com/codex

Two modes exist, and the safe one is the default for a reason. Ask-before-write means the agent proposes and waits. Full access means it can write, delete, and run commands without asking — that is a key, not a convenience toggle.

“PC access” for an agent scoped to one folder means exactly that: it can read and write inside that folder and run commands the session allows. It does not hand over your passwords, your other apps, your camera, or your whole disk — scope is the safety.

Remote control — an agent working while you’re away — multiplies what ask-before-write protects, because nobody is at the keyboard to catch a mistake live. Keep the safe default on for any remote session until you have a specific, reviewed reason to change it.

No database yet — just files and three prompts. A MEMORY.md index, a folder of daily notes, and a bootstrap prompt that makes the model read its own memory before it answers.

Copy-paste this at the start of a new session:

Before we start, read memory/MEMORY.md and today's file in memory/daily/
if it exists. Treat MEMORY.md as an index, not the full truth — if you need
detail, open the file it points to. Do not invent facts that are not in
these files. If nothing relevant exists yet, say so plainly and we will
start one.

Copy-paste this at the end of a work session:

Write today's entry in memory/daily/YYYY-MM-DD.md: what we did, what we
decided, and one open question for next time. Keep it factual. Then add
one line to memory/MEMORY.md only if today produced a fact worth finding
again later.

The full pack adds a third prompt — for correcting the model when it gets a fact wrong — and the exact file-creation steps for an agent.

GLAMMBRAIN is a set of local programs that makes a chosen folder of text files searchable after a chat ends. The files stay readable. The v4 feed builds one search aid for meaning and another for exact words; its search command combines matching passages and returns the filename behind each one. Separate jobs maintain the optional relationship map, summaries, journals, and snapshots. It is a real installation with ongoing refreshes, not a switch inside the AI model.

A question comes in — what happens next

1. First, you choose a configured folder of supported text files. A chat becomes a durable file only when it is written there — directly or by the session-summary job. The feed then makes that file searchable.
2. The feed splits long files into overlapping passages. It puts a numerical copy in Qdrant for meaning search and the actual words in BM25 for exact search.
3. A person or coding agent passes the question to unified-search.py. The v4 command checks those two indexes; it does not check the separate Neo4j graph.
4. The command merges both result lists. If the optional reranker is enabled and reachable, it reads the candidates again and adjusts their order.
5. The search returns evidence passages and source filenames. It does not write the final prose answer; the person or AI using the search decides what to say from that evidence.

READABLE FILES

The notes and other text you choose are the durable record. A person can open, correct, move, or remove them.

QDRANT · MEANING

Keeps numerical copies of passages with their source paths so different wording can still lead to a related idea.

BM25 · EXACT WORDS

Keeps the actual words from those passages so names, dates, codes, and quoted phrases remain findable.

NEO4J · OPTIONAL MAP

Separate jobs build file, folder, reference, and optional entity relationships. The normal v4 search does not consult this map.

NIGHT FILES

Session summaries and dream journals are written as Markdown that a later feed can make searchable. Proposals stay in a separate JSON file. If safe apply explicitly accepts an eligible deterministic patch, it appends that patch to a Markdown memory file the feed can index.

You do not need any of what follows to finish the trail: the six steps above already leave you with a working agent. If you want to install the GLAMMBRAIN memory on your own machine, the full technical procedure is one click away.

tar xzf glammbrain-20260827-full-v4.tar.gz && cd cockpit-20260827
cat brain-engine/AGENT-PACK.md

This station also has three short packs: the plain map, the standalone install work order, and the Mac guided-beta override. After downloading the archive, its own brain-engine/AGENT-PACK.md is authoritative for Linux/WSL; Mac users must keep the Mac override beside it.