I kept running into the same problem: ChatGPT would help me think through an architecture, Claude Code would help me implement it, Cursor or Windsurf would touch the repo later, and every handoff would lose context. Not just “what files exist,” but the stuff that actually matters: why we chose one approach, what we already rejected, what the project conventions are, what setup detail will bite later, and what the agent learned last time. So I built mem-port : a local MCP server that gives A...