(I've a recently acquired strong opinion on MCP, beware)

MCP seems to be the idea to expose literally any data source to AI agents as context. E.g, in work/business context: a database, source code, a kanban board; in private context: your photo library, notes, etc. This is good and makes sense.

As a technical "standard" or "protocol" (as it claims to be), it's a total mess, read more here: https://modelcontextprotocol.io. Though I guess they are inventing it as they go.

Nah, that's still considered an ephemeral user session. We're talking about what the service does with what it learns from the failures (and successes) marked by user feedback into the chat, or the app's buttons.