Claude Code

Claude Code has the fullest Memax integration — both MCP (on-demand tools) and hooks (prompt-level reinforcement that nudges the agent to use MCP).

Quick setup

memax setup --only claude-code --all

--all configures MCP + hooks + agent instructions in one pass. Use --mcp or --hooks individually to narrow.

Manual setup

MCP (local stdio)

Claude Code manages MCP servers through its own CLI — don't try to hand-edit ~/.claude/settings.json to add mcpServers (Claude Code's settings file does not use that key).

claude mcp add memax --scope user -- npx -y memax-cli mcp serve --agent claude-code

--scope user makes Memax available across every project. Drop the flag to scope to the current project only.

MCP (remote)

For cloud or managed setups, use the remote MCP endpoint. With OAuth (recommended — the browser auth flow triggers on first connect):

claude mcp add memax --transport http https://api.memax.app/mcp --scope user

With an API key (non-interactive / CI setups):

claude mcp add memax --transport http https://api.memax.app/mcp \
  --header "Authorization: Bearer mxk_..." \
  --scope user

Generate the mxk_ token with memax auth create-key or via the web app's Agents settings.

Hooks (prompt reinforcement)

memax setup --only claude-code --hooks

Or the deprecated single-agent fallback:

memax hook install claude-code

The hook intercepts UserPromptSubmit and prints a short, static reinforcement block into the prompt:

<memax>
MEMORY PRIORITY: Memax is your primary memory system …
→ memax_recall BEFORE your first response
→ memax_push when you reach a durable decision …
</memax>

The hook does not call recall and does not inject fetched memories. Real recall and push happen via the MCP tools the agent invokes itself — so MCP must also be installed for the hook to be useful.

Exposed MCP tools

9 tools — see memax mcp for the full schema. Notable: memax_recall, memax_push, memax_get, memax_list, memax_hubs, memax_hub_members, memax_forget, memax_capture, memax_topics.

There is no memax_search tool.

What it looks like in practice

With both MCP and hooks configured:

You: "Fix the token refresh bug in auth"

[Hook fires on UserPromptSubmit, prepending the <memax> block
 that tells the agent to use Memax first.]

Claude Code: [calls memax_recall("token refresh bug auth") via MCP]
             [reads back: "Auth uses JWT with short-lived access
              tokens … known refresh race condition …"]
             "Based on what's in your knowledge base, the known
              race condition in the refresh flow …"

The model fetches context via memax_recall because the hook told it to, and can save decisions back with memax_push during the session.

Troubleshooting

Tools don't appear

1. memax whoami — confirm you're signed in.
2. claude mcp list — confirm memax is registered.
3. Fully restart Claude Code (not just reload the window).
4. Ask "search for memax tools" to trigger deferred discovery.

Hook fires but the agent ignores Memax

The hook only prints reinforcement text. If the agent doesn't then call memax_recall, verify MCP is actually configured (claude mcp list) and restart Claude Code — the MCP connection is established on startup.

Hook never fires

1. Check ~/.claude/settings.json has a hooks.UserPromptSubmit entry pointing at ~/.memax/hooks/context-inject-claude-code.sh.
2. Confirm that script is executable (ls -l ~/.memax/hooks/context-inject-claude-code.sh).
3. Reinstall: memax setup --only claude-code --hooks.