memax mcp

Run the local (stdio) Memax MCP server. Gives tool-calling agents direct access to your hub.

Start a local MCP (Model Context Protocol) server over stdio. Any MCP-compatible client can connect to it and get access to Memax tools.

For most setups, you don't run this command directly — agents invoke it via their MCP config (see Integrations).

Usage

memax mcp serve [--agent <name>]

Flag	Description
`--agent <name>`	Tag the agent's identity in audit logs (`claude-code`, `cursor`, etc.).

Configuration

An agent configures the MCP server in its JSON config:

{
  "mcpServers": {
    "memax": {
      "command": "npx",
      "args": ["-y", "memax-cli", "mcp", "serve"]
    }
  }
}

For agent-specific setup, see the Integrations guides.

The CLI's local MCP server and the remote MCP server (at https://api.memax.app/mcp) expose the same tools. Use whichever fits your setup; remote is lower-friction for cloud deployments, local is better for offline / stdio-only agents.

Exposed tools

The MCP server exposes 9 tools. All parameter names and types below come directly from the server's inputSchema.

`memax_recall`

Semantic search across the user's hubs. Use it for project context, team processes, or past decisions.

Parameter	Type	Required	Description
`query`	string	yes	Natural language search query.
`limit`	number		Max results (default 10).
`topic_id`	string		Restrict to memories in this topic (UUID).
`hub_id`	string		Boost ranking for this hub (ID, slug, or `personal`).
`project_context`	object		`{ repo, project }` for relevance boosting.

`memax_push`

Save new knowledge. The server classifies it automatically.

Parameter	Type	Required	Description
`content`	string	yes	The content to save.
`title`	string		Auto-generated if omitted.
`hint`	string		Context hint (e.g. `"my resume"`) — improves classification.
`tags`	string[]		Tags.
`initiation_type`	string		`human_direct`, `human_requested_agent`, `agent_proactive`, `agent_automatic`, `import`, or `unknown`.
`project_context`	object		`{ repo, project, branch }`. Auto-detected by the CLI.
`hub_id`	string		Target hub (required for team hubs).
`hub_reason`	string		Why it belongs in the shared hub — required when `hub_id` is a team hub.

`memax_get`

Read the full content of a memory by ID — use this after memax_recall to read a result in full.

Parameter	Type	Required
`id`	string	yes

`memax_list`

List memories with pagination. Returns cursor for the next page.

Same input schema as the CLI's memax list: sort, limit, cursor, hub_id, topic_id.

`memax_hubs`

List hubs the current user can access. Returns hub IDs, slugs, roles, and memory counts. No input parameters.

`memax_hub_members`

List members of a hub.

Parameter	Type	Required	Description
`hub_id`	string	yes	Hub ID, slug, or `"personal"`.

`memax_forget`

Delete a memory by ID.

Parameter	Type	Required
`id`	string	yes

`memax_capture`

Save structured session context — summary, decisions, and learnings from a work session. Each decision / learning is extracted into its own searchable memory.

This is not a URL capture — it's for end-of-session context persistence from the agent's perspective.

Parameter	Type	Required	Description
`summary`	string	yes	Short summary of what was accomplished.
`decisions`	string[]		Key decisions made (one entry per decision).
`learnings`	string[]		Things learned during the session.

`memax_topics`

Browse the user's topic tree. Without topic_id: full tree with memory counts. With topic_id: memories in that topic.

Parameter	Type	Required	Description
`topic_id`	string		Topic ID. Omit for the full tree.
`hub_id`	string		Scope topics to a specific hub.

Tool parity

The tools listed above are identical between:

Local (stdio) MCP — memax mcp serve (this command).
Remote MCP — https://api.memax.app/mcp (hosted).

If you add or change tools in one, update the other — this is enforced as a contract in the repo.

On this page