ChatGPT

ChatGPT connects to AutoMem through a remote MCP sidecar — a separately deployed HTTP service that translates MCP protocol requests to AutoMem API calls. This is required because ChatGPT is cloud-based and cannot spawn local processes.

Note

ChatGPT requires URL-based token authentication for custom MCP connectors. Header-based auth is not supported by ChatGPT’s connector configuration UI.

---

Architecture

graph TB
    subgraph "Remote MCP (HTTP/SSE Transport)"
        CGP["ChatGPT Developer Mode"]
        SIDECAR["MCP Sidecar (separate deployment) — POST /mcp / GET /mcp/sse"]
        AUTOMEM["AutoMem API :8001"]

        CGP -->|"HTTP POST"| SIDECAR
        SIDECAR -->|"HTTP"| AUTOMEM
    end

• Local MCP (Claude Desktop, Cursor, Claude Code): AI platform → stdio → MCP server → AutoMem API
• Remote MCP (ChatGPT, Claude.ai, ElevenLabs): AI platform → HTTPS → MCP sidecar → AutoMem API

The MCP sidecar is part of the AutoMem server repository (mcp-sse-server/), deployed separately from your local tools.

---

Prerequisites

1. A deployed AutoMem service (see deployment guide)
2. The MCP sidecar deployed and accessible over HTTPS
3. A ChatGPT Plus, Pro, Team, or Enterprise account (Developer Mode required)

---

Deploy the MCP Sidecar

The sidecar is included in the AutoMem Railway one-click template. For manual deployment:

Required environment variables:

Variable	Purpose
AUTOMEM_API_URL	AutoMem service URL (e.g., http://memory-service.railway.internal:8001)
AUTOMEM_API_TOKEN	API token for AutoMem service authentication
PORT	Server listen port (default: 8080)

Railway deployment:

1. Deploy via Railway one-click template (includes sidecar)
2. Go to mcp-sse-server service → Settings → Networking → Generate Domain
3. Note the public URL: https://your-mcp-bridge.up.railway.app

Caution

On Railway, use the internal DNS (http://memory-service.railway.internal:8001) for AUTOMEM_API_URL to avoid egress charges. The memory service must have PORT=8001 set — without it, Flask defaults to port 5000 and connections will fail.

Sidecar endpoints:

Endpoint	Method	Purpose
/mcp	POST	Streamable HTTP — full-duplex MCP (recommended)
/mcp/sse	GET	SSE stream — server→client events (legacy)
/mcp/messages?sessionId=<id>	POST	SSE client→server messages (legacy)
/health	GET	Service health probe

---

Connect ChatGPT

1. Open ChatGPT → Settings → Connectors → Advanced
2. Enable Developer Mode
3. Click + Add Server
4. Enter the server URL

Streamable HTTP (recommended):

https://your-mcp-bridge.up.railway.app/mcp?api_token=YOUR_AUTOMEM_TOKEN

SSE (fallback if Streamable HTTP not supported):

https://your-mcp-bridge.up.railway.app/mcp/sse?api_token=YOUR_AUTOMEM_TOKEN

5. Save the configuration

---

Available Memory Tools

All six AutoMem tools are available via remote MCP:

Tool	Description
store_memory	Store content with tags, importance, and metadata
recall_memory	Hybrid search (semantic + keyword + tags + time)
associate_memories	Create typed relationships between memories
update_memory	Modify existing memory fields
delete_memory	Permanently remove a memory
check_database_health	Check FalkorDB and Qdrant connection status

---

Verification

Test the connection by asking ChatGPT:

Check the health of the AutoMem service

Then test storing and recalling:

Store a memory: I prefer dark mode in all applications

What are my preferences?

---

Troubleshooting

ChatGPT cannot connect to the MCP endpoint

1. Verify the sidecar is running: curl https://your-mcp-bridge.up.railway.app/health
2. Check DNS resolves for your domain
3. Ensure HTTPS (port 443) is open — ChatGPT requires TLS
4. Verify TLS certificate is valid (Railway domains use wildcard certs automatically)

401 Unauthorized

1. Check AUTOMEM_API_TOKEN is set in the sidecar’s environment variables
2. Verify the token in the URL matches exactly (no extra spaces)
3. Ensure the token is URL-encoded if it contains special characters
4. Test token directly: curl -H "Authorization: Bearer $TOKEN" https://your-automem.up.railway.app/health

Memory service connection errors (502/504)

Symptom in sidecar logs:

[AutoMem] Fetch failed for http://memory-service.railway.internal:8001/memory:
Error: connect ECONNREFUSED fd12:ca03:42be:0:1000:50:1079:5b6c:8001

Cause	Solution
Missing PORT=8001 in memory service	Add PORT=8001 to memory service environment variables
Wrong internal hostname	Update to memory-service.railway.internal:8001
Service not running	Check Railway dashboard for deployment status

Variable resolution failures

Symptom: Logs show literal ${{...}} syntax:

[AutoMem] GET http://${{FalkorDB.RAILWAY_PRIVATE_DOMAIN}}:6379/health

Railway variable references (${{...}}) only work in template deployments. Use hardcoded values for manual service configuration.

SSE connection drops

If using SSE transport (legacy), the sidecar sends heartbeats every 20 seconds. Some proxies buffer SSE streams — ensure your proxy is configured for streaming responses.

Consider switching to Streamable HTTP (/mcp) instead of SSE (/mcp/sse) for better reliability.

---

Security Notes

• URL-based tokens (?api_token=) appear in server logs and browser history
• ChatGPT’s connector UI does not support custom headers, making URL-based auth unavoidable
• Use Railway’s private networking to keep the sidecar-to-AutoMem communication internal
• Rotate your AUTOMEM_API_TOKEN periodically

---

Related Platforms

Other cloud-based platforms using the same remote MCP sidecar:

• Claude.ai (Web) — supports header-based auth
• ElevenLabs — uses SSE transport with header auth