MCP agent memory

The hosted MCP server exposes memory, cognition, web search, composition, and autonomous task tools. It is mounted by FastAPI at /mcp and backed by cogix/integrations/mcp/server.py.

Connect

Use the hosted endpoint:

https://api.cognitivx.io/mcp

For API-key auth, send:

Authorization: Bearer icog_...

MCP clients that support OAuth discovery can use the built-in PKCE flow from /.well-known/oauth-authorization-server.

Tool habits

• Call recall before making claims about past decisions.
• Call remember after decisions, fixes, schema changes, production incidents, and user preferences.
• Prefer mode: "deep" for stable facts and mode: "shallow" for one-off session events.
• Use update instead of adding a second memory when the old memory is wrong.
• Use forget for bad records that should not be recalled again.

Agent profiles

Agents can register through /api/agents/register. The backend stores a slug, agent type, current task, and context summary. The talk route can use agent_slug and current_task so iCog distinguishes the current task from old exchange memories.

Long tasks

The MCP server also exposes agent_task and agent_status. agent_task launches an in-process background loop with access to selected memory tools and returns a task id. Poll with agent_status until it completes or fails.

Because this task store is in-process, it resets on backend restart. Treat it as an active-session helper, not a durable queue.