The MCP point is right, though tbh MCP is more like plumbing than memory. Execution-time context for tools and resources. The world model is a different thing entirely, it needs to persist across sessions, accumulate, actually be queryable.

In practice it's mostly RAG over structured artifacts. Process docs, decision logs, annotated code and so on. Conversation history works better than you'd expect as a starting point but gets noisy fast and I haven't seen a clean pruning strategy anywhere...

On the format question imo nobody really knows yet. Probably ends up as some kind of knowledge graph with typed nodes that MCP servers expose or so, but I haven't seen anyone build that cleanly. Most places are still doing RAG over PDFs so. That tells you where the friction is.