How to set up an MCP memory server for ChatGPT, Claude & Cursor
Set up an MCP memory server for supported clients. Compare authentication options, configure each tool, and verify context reads and writes.
An MCP memory server lets compatible AI clients access the same saved context. To share memory across Claude, Codex, Cursor, OpenClaw or Hermes, connect each client to one store, authenticate it, and confirm its memory tools actually run. A connected icon alone does not prove that anything was saved.
What the server does
MCP is a protocol for using external tools. A memory server exposes tools that save exchanges, retrieve relevant records and search past context. The server does not automatically inherit every conversation from the clients connected to it. It can retrieve the content that reached its store.
Vilix AI provides a hosted Streamable HTTP endpoint at https://api.vilix.ai/mcp. Its core conversation loop uses get_context to retrieve context and save_turn to persist the completed exchange. Search, projects and tasks, rules, reusable skills and agent inbox tools are also available. MCP tools/list is the current tool inventory.
Choose the authentication path for your client
OAuth is the browser sign-in path for clients such as Claude, Codex, Cursor and ChatGPT. Add the Vilix AI server URL in that client's settings, approve the connection and enable its tools in the conversation. Repeat this in each client, using the same Vilix AI account.
For headless clients such as OpenClaw or Hermes, create an API key in Vilix AI under Agents → Create API key. Configure the remote endpoint with Authorization: Bearer YOUR_VILIX_API_KEY. The API key and OAuth paths share the account's ownership and plan limits. Keep real keys out of chat transcripts, screenshots and committed configuration.
Client requirements checked on 10 September 2026
Claude web and desktop support custom connectors on Free, Pro, Max, Team and Enterprise. Free permits one custom connector. Team and Enterprise owners may need to add the connector before members authenticate. Mobile installation is in beta; use the web or desktop flow if it is missing.
Codex supports remote MCP through its desktop, CLI and IDE configuration. The configuration belongs to the host where it is installed. The CLI setup is codex mcp add vilix --url https://api.vilix.ai/mcp followed by codex mcp login vilix.
Cursor needs a server entry and authorization, followed by an Agent-mode conversation with the memory tools enabled. Configure Cursor on its desktop or CLI host.
OpenClaw and Hermes run on your configured computer or server. Use the client-specific guide for Bearer-header syntax and connection checks, then start a new agent session.
ChatGPT Developer mode requires Plus, Pro, Business, Enterprise or Edu on the web. Enable it under Settings → Security and login, then use Plugins → + to create an OAuth app. Workspace rules and write approvals still apply.
Gemini CLI and Gemini Spark are separate products with separate requirements. This is not blanket support for ordinary consumer Gemini chat. Spark requires eligible Google AI accounts and may ask for confirmation before saving memory. Gemini CLI eligibility changed in June 2026; check the current provider requirements.
Enable the memory workflow
Add Vilix AI's auto-memory instructions to the client's persistent instructions or your project rules. The published workflow asks the client to retrieve context before composing a reply, save the exact completed exchange, and reuse chat_id within the same conversation. Respect the client's approval prompts. If the tools fail or a save is declined, the assistant should say so rather than imply the conversation was stored.
Try a cross-tool handoff
Use a fictional project with no private information. In your first connected client, ask it to save: "Northstar is an invoicing app. The pricing page comes first. The launch date is 20 October." Confirm the save tool succeeded.
Open a new conversation in a second connected client and ask: "Retrieve Northstar from Vilix AI. What are the priority and launch date?" Inspect the retrieval result and the source conversation instead of accepting a plausible answer alone.
Next, save a correction in the first client: "Northstar's launch moved to 27 October." Ask the second client again. Finally, ask for something you never saved, such as the customer's phone number. A useful result should distinguish the saved facts, the correction and missing information. This is a suggested check, not a published performance benchmark or a claim that every client has been tested.
Troubleshooting
For missing context, check that both clients use the same account and that a save succeeded. For a 401, reconnect OAuth or replace a revoked or expired key. For daily_limit_reached, new saves are paused until the allowance resets at UTC midnight or a higher plan increases it; existing memory remains readable. For read_rate_limited, honor retry_after_seconds before retrying.
New accounts receive a seven-day Pro trial without a credit card. Without an upgrade, the account moves to Free. Existing history, recall and search stay accessible. Daily saving allowances and project-board limits differ by plan, and burst protection can briefly pace rapid reads.
Use the current Get started guide for exact client instructions and the MCP reference for transport and tool details. Configure one client first, verify a real save and recall, then add your next tool.