humaux-memory-mcp
v0.3.0
Published
Connect any MCP agent to your Humaux Memory — persistent, structured memory that every session shares. Self-updating, with resilient auto-reconnect across server redeploys.
Maintainers
Readme
humaux-memory-mcp
Connect any MCP-capable agent to your Humaux Memory — persistent, structured memory that every session shares. Your agent gains tools to store and recall memories, read your persona and skills, and browse the corroborated public commons.
This launcher delegates protocol bridging to mcp-remote (pre-filling the Humaux endpoint and your auth header) and adds two reliability layers:
- Self-update — on startup it checks npm for a newer version and, if found, transparently re-execs the latest one. You never have to manually bump the connector to get fixes; new releases reach every agent automatically. (Fail-open: any check failure just runs the current version.)
- Supervised reconnect — if the connection dies (most commonly right after a server redeploy) it detects the drop via health-probe / stalled-request / child-exit signals, respawns the bridge, and transparently replays the MCP handshake so your session survives without a manual reconnect.
Opt-outs: HUMAUX_NO_SELFUPDATE=1, HUMAUX_NO_SUPERVISE=1. Tunables: HUMAUX_HEALTH, HUMAUX_HEALTH_INTERVAL_MS, HUMAUX_REQUEST_TIMEOUT_MS.
Usage
Get your token at https://humaux.com → Connect, then:
npx -y humaux-memory-mcp --token mos_xxxor set it in the environment:
HUMAUX_TOKEN=mos_xxx npx -y humaux-memory-mcpAdd it to your agent
Claude Desktop — claude_desktop_config.json:
{
"mcpServers": {
"humaux-memory": {
"command": "npx",
"args": ["-y", "humaux-memory-mcp", "--token", "mos_xxx"]
}
}
}Cursor — ~/.cursor/mcp.json, Cline, Windsurf, Zed and any other stdio MCP client use the same command / args shape.
For HTTP-native clients (e.g. Claude Code), you can also connect directly:
claude mcp add --transport http humaux-memory https://humaux.com/memory/mcp \
--header "Authorization: Bearer mos_xxx"What your agent gains
No prompt required — MCP is self-describing. On connect, your agent sees these tools and knows when to use them:
memory_store— save a decision, fact, or preference worth keepingmemory_search— recall the right memory (dense + sparse + rerank)memory_meta— pull persona, skills, related memories & contradictionscode_search/doc_search— query your code graph and documentspublic_browse— read the corroborated public commons
License
MIT
