David decided: Rejected pulse-based MCP cadence in favor of memory-layer reframing

After the agent proposed two mechanisms — a server-side granular_pulse tool the model is required to call, and a repositioning of granular as Claude's memory layer — David probed the memory-layer approach:

but RE A, wouldnt it have to load the granular connector to access the MCP tools first?

This question surfaced the bootstrap problem: repositioning granular as a memory tool only works post-install and post-activation; it doesn't solve the problem for users who haven't yet enabled the connector. David's follow-up then challenged the agent's claim that Claude Code connectors are always-on, which prompted the research task that confirmed per-chat activation on claude.ai web. The distinction mattered for how to design the cadence mechanism.