Skip to main content

Connect an agent

This is the standard setup flow for Claude, Codex, Hermes, OpenClaw, Gemini, Cursor, VS Code, and direct REST/Gateway clients.

1. Choose the agent​

Open the UnifiedMemory app and go to Connections. Pick the agent you are configuring. The registry decides the default source app, capabilities, install command, and verification steps for that agent.

Default production values:

SettingValue
Edge base URLhttps://edge-api.jithendranara.dev
MCP URLhttps://edge-api.jithendranara.dev/mcp
Containerjeethendra
Agent scopeown

2. Issue a scoped key​

Use a scoped agent key, not an admin key. The key should be bound to the agent and include only the capabilities the integration needs.

Scoped key rules:

  • Effective identity comes from key metadata, not request body fields.
  • agent_scope=own is the default.
  • Working and session memory stay private to the same agent/app/run.
  • Shared owner facts require promotion through trust/review gates.
  • Admin and forensic visibility are separate operator routes, not prompt context.

3. Install the client configuration​

Use the copy-ready command or JSON config from the selected integration page:

4. Verify the key​

Verification should check the selected key, not just your browser session. A real integration check confirms:

  • the key is active;
  • the key can access jeethendra;
  • the key is bound to the expected source_app and agent_id;
  • advertised capabilities match the configured tools;
  • a simple tool call succeeds.

5. Verify memory behavior​

Run one recall or context request, then one writeback request with a disposable fact. Confirm:

  • own-agent memory is visible to the same agent;
  • another agent's working memory is not visible;
  • approved canonical owner facts can be shared;
  • Courtroom can inspect important returned memories;
  • chat crumbs and assistant dumps are rejected or quarantined.

6. Read Courtroom​

For any memory that affects an answer, open the Courtroom link. You should see the deterministic verdict first, then cached/generated MiniMax judgment when available.

Next steps​