Detailed Analysis
A creator-led experiment addresses a gap in Claude's current product design: the platform's memory and connector architecture is fundamentally single-player, even as teams increasingly want to collaborate through it. The article stems from community polling showing nearly 70% of a given user base is trying to use Claude collaboratively, yet almost none have found a clean way to do so. The diagnosis is structural—Claude's native memory ties to an individual account, third-party memory frameworks store data locally on one machine, and MCP (Model Context Protocol) connections authenticate per-user. None of these were built with shared team context in mind, which the author validates by pointing to Y Combinator's Request for Startups list, which explicitly names "team infrastructure for agents" as an unsolved problem worth funding.
The proposed workaround stitches together three existing tools rather than waiting for a purpose-built solution: Notion serves as the shared, permissioned "source of truth" for team documents (brand guidelines, voice profiles, contacts); Supermemory.ai acts as a queryable shared working memory that captures what the team learns as they operate, so one person's context and decisions become visible to teammates; and Claude's desktop app remains the interface layer, chosen specifically for its non-technical, friendly UI. Critically, no custom software is built—each team member connects to Notion via their own MCP login, meaning Claude can only read and write to pages that person already has permission to access in Notion. This means the permission system isn't reinvented; it's inherited wholesale from Notion's existing access controls, which the author contrasts favorably against messier alternatives like manually syncing CLAUDE.md files and project folders across machines, or using GitHub, which is too technical for typical business users.
This matters because it exposes a broader tension in the current wave of AI agent tooling: individual productivity gains from tools like Claude have outpaced the infrastructure needed to make those gains organizational rather than personal. As AI assistants become embedded in daily workflows, the unit of value creation is shifting from "my chat history and memory" to "our shared institutional knowledge," and vendors have been slow to catch up. Anthropic's decision to build Claude around MCP as an open, extensible connector standard is what makes this kind of workaround possible at all—third parties can plug enterprise tools like Notion into Claude without Anthropic needing to natively solve team memory itself. That architectural choice is proving to be a meaningful differentiator, letting the ecosystem route around limitations faster than any single vendor roadmap could.
More broadly, this reflects a maturing pattern in enterprise AI adoption: rather than waiting for vendors to ship "Claude for Teams" with built-in shared memory, sophisticated users are assembling composable stacks from existing SaaS tools (Notion, Supermemory, MCP connectors) to approximate that functionality today. It's a signal to Anthropic and competitors alike that shared, permissioned, queryable memory is a top unmet need—one that will likely get formalized into native product features as the market validates demand through exactly these kinds of grassroots implementations. The explicit acknowledgment of limitations (an "80% solution for 20% of the work") also underscores that this remains a stopgap, not a permanent architecture, foreshadowing an inevitable wave of dedicated agent-teamwork infrastructure startups and enterprise features.
Read original article →