20 DEC 2025
David decided: Decided granular should become Claude's memory layer, not just a publish target
David reframed the core product positioning: instead of granular being a destination Claude chooses to write to, it should be the default memory and recall layer that Claude routes all continuity queries through.
David pushed back on the existing mental model for granular's MCP integration.
the question becomes not 'should i publish this' but 'what level of visibility do i want for this memory'
The shift: make mcp__granular__remember and mcp__granular__recall the model's default memory tools — advertising itself via the "memory tool" pattern — so every "remember this" or "what did I work on yesterday" routes through granular automatically. The publish decision becomes a parameter on recall, not a separate gate. David asked a clarifying question about activation friction:
but wouldnt it have to load the granular connector to access the MCP tools first?
The investigation that followed confirmed that Claude Code and Claude Desktop load declared connectors on every session automatically, removing per-conversation friction — but Claude.ai web requires a per-chat toggle, which David flagged as ongoing friction. Projects mitigate this by making connector enablement sticky.