Records the user's screen entirely locally and compiles activity into structured frames served over MCP, giving AI agents searchable episodic memory without cloud processing or LLM intermediation.