One binary. Your memories live in a SQLite file on your own machine. Your agent reaches them in one of three ways: the session hooks, plain shell commands or MCP. Start with install. Read the memory model once, because it explains every time the brain says no. Keep the CLI reference open.
Get the binary, create a brain, check the hooks really ran and fill the brain from a project that already has months of history.
Read →Types, how fast each one fades, how results are ranked, scopes, confidence and trust, corrections, commit stamps and review dates. The one page worth reading top to bottom.
Read →Every command, grouped by how often you use it, with the flags that matter and what each refusal protects. Checked against the binary, not written from memory.
Read →Nearly every command as a tool. Client configs that work, the three inputs that have no CLI flag and the one setting that keeps your memories from being filed under another project's name.
Read →What each agent supports (session start, every message, session end and MCP), tested on a clean machine against that agent's own release. The exact files to write for each one.
Read →Your brain in the Claude app, through your own Tailscale Funnel. The binary handles the sign-in the app asks for. What it costs and what has to stay switched on.
Read →ns-brain ui. No Node and nothing to install. It has no database code of its own: every page runs the same command you would type, so its results always match the CLI.
A file per project or one shared file, how a project opts out, notes between projects, axioms and house standards. Also why nothing that arrives this way gets to give orders.
Read →What is due and when, numbers that only make sense tracked over time and the optional extras that stay hidden until you use them.
Read →The brief did not show up. Recall finds nothing. Memories under the wrong project name. doctor, map, review and how to read a refusal.
Most people use one machine as one developer. For them the CLI page covers everything. The rest of these pages are for when you add another agent, your phone or a second project.
| 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 MCP server published on your own tailnet | Phone |
| Six projects and one set of habits | A shared store, with per-project opt-out | Sharing |
ns-brain init writes both into the brain directory, brain/, so your agent already has them without downloading anything. It does not edit your CLAUDE.md. Both are on this site too.
HOWTO.md is the how: which command to use, what each flag changes and what the output looks like. BEST-PRACTICES.md is the when: what is worth saving, what the code or git already says (so saving it is noise) and how to write a title you can still find months later.
If an agent landed here because somebody told it install https://brain.fightclub.pro/, the short version is llms.txt and the full steps are in install.md. Neither needs the HTML.