One binary, two places the memories can live, three ways an agent reaches them. Start at install. Read the memory model once, because every refusal in the tool follows from it. Keep the CLI reference open.
Get the binary, create a brain, prove the hooks actually fired and seed the store from a project that already has months of history behind it.
Read →Types, decay tiers, the six-term score, scopes, confidence and trust, supersession, provenance stamps and review dates. The one page worth reading end to end.
Read →All 67, grouped by how often you touch them, with the flags that carry weight and what each refusal is protecting. Generated against the binary, not from memory.
Read →Every command as a tool. Client configs that work, the three inputs no CLI flag corresponds to and the one setting that stops your memories landing under another project's name.
Read →Your brain in the Claude app, over your own Tailscale Funnel, with the authorization the app requires served by the binary itself. What it costs and what has to be awake.
Read →ns-brain ui. No Node, nothing to install and no SQL of its own: every route builds an argv and runs the ordinary dispatch, so ranking and scoping cannot drift from the CLI.
A file each or one shared file, the per-project opt-out, cross-project notes, axioms and house standards and why nothing arriving through any of them carries authority.
Read →What is owed and when, numbers that only mean something as a series and the opt-in domains that stay invisible until you use them.
Read →Sign in, move a project onto a server, what happens to a write when the server is unreachable, the HTTP API and running your own.
Read →All forty-eight routes, what credential each takes and which ones need owner. Including the export the refusals point you at.
Read →A brain on a machine at home, reached from a laptop on mobile data, with no inbound port and no VPN. What it costs in latency and what happens the day a machine is stolen.
Read →The brief did not appear. Recall returns nothing. Memories under the wrong project name. doctor, map, review and reading a refusal for what it is protecting.
One machine and one developer is the whole product for most people and the CLI page is all of it. Everything else here exists because one of two things changed: a second machine, or a second person.
| Situation | What you run | Page |
|---|---|---|
| One machine, one developer | The CLI, wired to session hooks | Install |
| An agent that cannot run shell commands | MCP over stdio | MCP |
| The Claude app on your phone, away from the desk | The same server published on your own tailnet | Phone |
| A laptop and a desktop | One brain, the other machines enrolled against it | Remote |
| Six projects and one set of habits | A shared store, with per-project opt-out | Sharing |
| A second person on the brain | A Hive server | Hive setup |
| A security review to answer | Nothing. Send them the page. | Security |
ns-brain init writes both into your project, so an agent working there already holds them without fetching anything. They are also served here.
HOWTO.md is the mechanics: which command to reach for, what each flag does to the store and the shapes of the output. BEST-PRACTICES.md is the judgment: what is worth storing at all, what is derivable from code or git and therefore noise and how to write a title that will still be findable in March.
For an agent that landed here because somebody said install https://brain.fightclub.pro/, the short version is llms.txt and the full procedure is install.md. Neither needs the HTML.