> ## Documentation Index
> Fetch the complete documentation index at: https://manager.harmix.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Quickstart

> Connect Memory MCP to your agent in minutes: generate a key, paste config, and run retrieve_memory.

Get from zero to a successful `retrieve_memory` call. Choose the path that matches your client.

## Prerequisites

Before you begin:

* A PAM account with at least one connected memory source
* Memory readiness in **Ready** state (see Memory MCP in the PAM app)
* An MCP-compatible client

## Desktop agents (Cursor, Claude Code, VS Code)

<Steps>
  <Step title="Generate an agent API key">
    Open **[Setup](https://pam.harmix.ai/setup)**, choose **Desktop app or script**, and click **Generate key**.

    Copy the full key when shown — you cannot view the secret again after leaving the page. Keys use the format `pam_mkey_<key>`.
  </Step>

  <Step title="Add the MCP server to your client">
    Paste this configuration into your MCP client settings. Replace the placeholder with your full key.

    ```json theme={null}
    {
      "mcpServers": {
        "pam_memory": {
          "type": "http",
          "url": "https://api.pam.harmix.ai/v1/mcp/memory",
          "headers": {
            "Authorization": "pam_mkey_<key>"
          }
        }
      }
    }
    ```

    See [Client setup → Desktop agents → Cursor](/docs/client-setup#desktop-agents) (or Claude Code / VS Code tabs) for platform-specific steps.
  </Step>

  <Step title="Add agent instructions">
    Add this snippet to your agent system prompt so it knows when to retrieve memory:

    ```text theme={null}
    You have access to PAM Memory through the retrieve_memory MCP tool.

    Use it when a request may depend on company-specific context:
    people, projects, decisions, processes, priorities, history, customers,
    documents, meetings, or internal terminology.

    Prefer retrieving memory before making assumptions.

    If retrieved context is partial or uncertain, say so and ask a focused follow-up.
    ```
  </Step>

  <Step title="Run a hello world prompt">
    Ask your agent:

    ```text theme={null}
    What do we know about our current project priorities, and who is responsible for follow-up on each?
    ```

    Your agent should call `retrieve_memory` with your prompt. A successful call marks your key as **MCP verified** on [Setup](https://pam.harmix.ai/setup).
  </Step>
</Steps>

## Browser LLM connectors (ChatGPT, Claude web)

<Steps>
  <Step title="Copy your MCP server URL">
    Open **[Setup](https://pam.harmix.ai/setup)**, choose **Web browser LLM**, and copy the **MCP server URL**.
  </Step>

  <Step title="Add a custom connector">
    In ChatGPT or Claude web, add a custom MCP connector and paste the URL. No OAuth client ID is required — the connector registers via DCR automatically.

    Follow [Client setup → Browser LLM → ChatGPT](/docs/client-setup#browser-llm-connectors), [Claude web](/docs/client-setup#browser-llm-connectors), or [Perplexity](/docs/client-setup#browser-llm-connectors). No OAuth client ID is required when discovery is enabled.
  </Step>

  <Step title="Approve access in PAM">
    Sign in to PAM if prompted, then click **Approve** on the consent screen (e.g. *"chatgpt.com will be able to read your PAM memory."*).

    **Cancel** returns you to the browser LLM without granting access.
  </Step>

  <Step title="Run a hello world prompt">
    Ask your browser LLM the same hello world question as above and confirm the call appears in **[Request Log](https://pam.harmix.ai/request-log)**.
  </Step>
</Steps>

<Check>
  Setup complete when your agent returns company context from a successful `retrieve_memory` call.
</Check>

<Tip>
  Need the full protocol reference? See [MCP reference](/docs/mcp-reference) and [retrieve\_memory](/docs/retrieve-memory).
</Tip>