For a year, I've been writing notes in a system called IWE. It's a small command-line tool that treats markdown files as a graph — notes link to each other, and you can navigate, search, and restructure without directly touching the files. It's my working memory layer as a strategic advisor.
This week, an MCP server was added to it. And it's the first time I've had a real experience of sharing memory with a machine.
What MCP is, in short
Model Context Protocol is Anthropic's attempt to standardize how AI models talk to external systems. Think of it like USB-C for LLMs: one port, many tools. An MCP server exposes a set of functions, and the model can call them like any other function.
It sounds technical, and it is. But the consequence isn't technical. The consequence is that the model is no longer a conversational partner that forgets everything when the window is closed. It can write to persistent storage that I also write to. And read what I've written since last time.
Thirteen tools, one shared space
The IWE server gives the model thirteen functions: search, retrieve with context, show hierarchy, create, update, delete, rename, extract paragraphs into new notes, flatten block references. These are the operations a knowledge worker performs on their note-taking system. Now the model can do them too.
It sounds small. It's not.
Before MCP, I had two options: copy my notes into the conversation manually, or let the model work blind. Now it can look up my knowledge graph itself, retrieve the right note with one layer of parents and two layers of children, and answer my question with sources from my own thinking.
When I write a blog post, it retrieves the tone from brand-and-message, the arguments from relevant source notes, and discovered connections I had forgotten myself. When I propose a package to a client, it pulls up history from client meetings I had noted down three months ago.
The asymmetry that disappears
The interesting part isn't that the AI can now work with my data. It's that it can maintain it.
When I discover a gap in my knowledge base — a note that should be written, a connection that's missing — I can ask the model to create it. It writes the note, adds it to the relevant MOC, and links it into the graph. The next time I sit down to type, the gap is closed.
It's a new kind of asymmetry. Previously, I was the producer, the model the consumer. Now we are both. My note-taking system grows faster than I can write myself.
This is also why I call it shared memory and not AI assistance. An assistant fetches coffee. Shared memory changes how you think, because you know you can externalize parts of your cognition and get them back structured.
What's not hype
I've been on stage giving talks about AI for a few years. I've seen all the fiascos — time and again, companies have been tasked with "integrating AI" and ended up with a ChatGPT subscription and a PowerPoint.
MCP is not like that. It's not a model getting smarter, not a new interface, not another chatbot. It's infrastructure. Plain, unsexy infrastructure that makes it possible to connect models to what we're already working with.
Infrastructure is what creates real productivity boosts. Models getting 10% better at benchmarks don't do that. You don't feel that difference in your daily life. The difference between a model that can read your notes and one that can't — you feel that immediately.
What it means for SMEs
I advise small and medium-sized businesses on AI. This is relevant for them, but not in the way many think.
What's relevant isn't that they have to set up IWE themselves. It's that the MCP architecture — a model connected to the company's own data store via a standardized protocol — is what will become the norm. The most valuable AI implementations in the next two years aren't new models. They are models connected to the CRM, the document archive, the accounting system.
And here's the point: if your company doesn't have its data in a format that can be exposed via an MCP server, you're not AI-ready. No matter how many ChatGPT licenses you buy.
Data order is the new prerequisite. MCP makes it clear why.
The little experiment
I have a little test. When I meet a company that says they are working on AI, I ask: "If you had to connect an AI model to your customer history today, how many hours would it take you to pinpoint where it is?"
The answer is usually weeks. Sometimes they don't know themselves. It's not an AI problem. It's a data foundation problem. And AI makes that problem visible in a way nothing else has.
I'm sitting here with my little markdown library connected to a language model via a protocol that didn't exist a year ago. It's not impressive because the technology is advanced. It's impressive because my homemade note-taking system is now more AI-ready than most companies' CRMs.
That's the distance boards should consider. Not tomorrow. Today.