RxAi AMP · Quickstart ·
You need a GitHub account and about ten minutes. One command creates a private memory repository that you own; one more connects each agent you use. From then on every session starts with the memories that match its task, and a session that committed work ends with one checkpoint asking what is worth keeping.
| Tool | Check | Notes |
|---|---|---|
| Git 2.x | git --version | |
| Node.js 22 | node --version | npm comes with it |
| GitHub CLI 2.x | gh auth status | brew install gh, then gh auth login |
| GitHub MCP server | github-mcp-server --version | brew install github-mcp-server — the wizard offers to install it |
| A GitHub account | Your memories live in a private repository on it |
Working with others? Each person connects with their own gh auth login. On a repository in a personal account, collaborators cannot use fine-grained tokens; move it into an organization if members need one-repository tokens (README Step 5).
The setup wizard runs on macOS and Linux. On Windows, follow the step-by-step guide in the README.
git clone https://github.com/RxAi-Aus/amp.git my-agent-memory
cd my-agent-memory
npm install
npm run setup -- --dry-run # see every step first; changes nothing
npm run setup
The wizard creates a private repository for your memories (it never pushes to the template), lets GitHub Actions write the index, adds the issue labels the protocol uses, connects your agents with your own gh login (a fine-grained token only for machines nobody logs in to), installs and registers GitHub’s official MCP server without writing any token into its config, and finishes by compiling the index and posting a real test issue. If a step fails it prints the manual command for it; running the wizard again resumes, and it never resets a repository that is already live.
Hooks decide when memory moves; the bundled rxai-amp skill tells the agent how to read and write it. Run the line for each agent you use. L2 means the hooks deliver memory; L1 means the agent is told the protocol and reads memory itself. The overview explains the levels.
| Agent | Command | What you get |
|---|---|---|
| Claude Code | npm run hooks:install:claude | L2: pointers at session start, summaries matched to each prompt, a checkpoint when work was committed, and the /amp command |
| Codex | npm run hooks:install:codex | L2: memories at session start and the checkpoint. Open /hooks in Codex once and trust the new hooks |
| agy | npm run hooks:install:agy | L2: memories before the first reply and the checkpoint |
| OpenClaw | npm run hooks:install:openclaw | L1: the protocol digest in its workspace AGENTS.md |
| Hermes | npm run hooks:install:hermes | L1: the protocol digest in ~/.hermes/SOUL.md, loaded even outside a project folder |
| Any repository you commit from | npm run hooks:install:capture -- /path/to/repo | A capture reminder after each git commit, for agents without hooks |
Add -- --dry-run to any installer to see what it will change before it changes anything.
=== RxAi AMP shared memory — RECALL. If the block is missing, the hooks are not installed or not trusted, and the agent says so instead of searching the index by hand.npm run setup:verify re-runs the whole checklist whenever you like./amp status shows which memory repository and agent name it resolved./amp update <what to remember> posts a memory issue in the protocol’s title format, or an Outcome: comment on the thread it belongs to. On its own, /amp update stores a summary of the session.gh issue view away.Outcome: success or Outcome: failure on the memories you used. The next index compile raises or lowers their weight, so what worked rises.