A consultant showed me his setup last week. He was proud of it — and I understand why. Twelve months of work. Claude wired into Obsidian through a custom MCP server. A vault of 4,000 notes, all interlinked, tagged, feeding back into prompts he'd refined over hundreds of iterations. Memory files. Context windows. A hand-built brain for his business.
He asked me what I thought. I told him the truth, which is that he'd built something genuinely impressive — and that most of it would be dead weight inside six months.
That wasn't what he wanted to hear. But it needs saying, because this isn't one consultant. It's a whole cottage industry of wannabe experts peddling Claude-and-Obsidian setups like it's cold fusion. And the founders buying it have no idea what they're actually being sold.
The Workaround Disguised as Architecture
Today's frontier models have real limitations. Context windows fill up. Memory across sessions is patchy. The model forgets what you told it last Tuesday. So a particular breed of consultant has appeared, selling scaffolding to compensate — note systems, retrieval pipelines, prompt libraries, vector indexes, custom MCPs, the whole stack. Obsidian sits at the centre of most of these because it's local, it's flexible, and it gives the buyer the comforting illusion that they've solved memory.
They haven't. They've paid five figures for a workaround to a temporary problem.
And the people building the most scaffolding are usually the ones who couldn't tell you what they're scaffolding around if you put a gun to their head. Every line of glue code, every custom integration, every clever indexing scheme is a confession that they don't actually understand the tool — they're just stacking complexity on top of complexity until it looks impressive enough to invoice for.
The tools you build to compensate for what the model can't do today will be the first things you tear out tomorrow.
Why This Particular Stack Won't Survive
I'm not picking on Obsidian. It's a brilliant note-taking app, and if you want to take notes, knock yourself out. I'm picking on the people selling it as a memory system to founders who don't know any better.
Three reasons it falls apart.
One: they're betting against the model. Claude's memory is getting better every quarter. Context windows are getting longer. Agentic memory — actual, persistent, structured memory the model maintains for itself — is months away, not years. Every consultant selling a hand-wired memory pipeline is selling you something Anthropic is about to ship for free. They know it. They're just hoping you don't.
Two: a markdown vault is not memory. It's a filing cabinet wearing a costume. Every time the model wants to use that information, it has to load it, parse it, retrieve it, summarise it, and re-inject it. That's not how a brain works. That's how a person works when they're using a brain incorrectly. Calling it "memory" is the same trick the crypto guys pulled when they called a database a "blockchain."
Three: you've been sold a glass cathedral that you have to keep dusting. Every new note category, every schema change, every model update breaks something downstream. The consultant who built it for you is gone or invoicing you again to fix it. Meanwhile, the founders who didn't fall for this are using their time to actually run businesses.
What This Actually Reveals
This is the part nobody wants to say. The complexity of someone's tool stack is usually inversely proportional to their ability to direct AI properly.
If you can clearly articulate what you want — in language the model understands, with the right context loaded in the right order — you don't need 4,000 interlinked notes and a custom MCP server. You need a sharp brief, the right model, and the discipline to say what you actually mean.
Most of the people building these elaborate Obsidian-Claude setups have never sat down and figured out how to talk to the model. They've abstracted their thinking into a system because their thinking isn't sharp enough to land in a single conversation. The tooling is a symptom. The disease is not knowing what good direction looks like.
And here's the kicker. The wannabe experts selling this stuff to founders are the worst offenders. They couldn't write a clean prompt to save their lives, so they've built an entire methodology around hiding that fact. Diagrams. Configs. Repository structures. It looks like expertise. It's actually the AI equivalent of a magician using both hands to distract you from the trick — which, in this case, is that there is no trick. They just don't know what they're doing.
If you can't tell me what your memory architecture is supposed to do in one sentence, you don't have a memory architecture. You have a folder with extra steps.
The Compounding Problem
Every hour spent maintaining a workaround is an hour not spent learning to direct the model properly. And those two skills compound very differently.
The operator who learned to write tight, structured prompts in 2024 is ten times more effective at directing agents now. The operator who paid for a custom Obsidian-Claude memory pipeline in 2024 is now paying again to migrate it, patch it, or rebuild it for the new model. One built skill. The other built sand.
So…
If you've bought one of these setups, I'm sorry. You were sold a story.
If you're about to buy one, don't.
If you're selling them, knock it off. The model is coming for your business model and it's going to be embarrassing when it happens.
And if you're one of the wannabe experts peddling Claude-and-Obsidian like it's cold fusion — you might want to start learning what you actually do for a living, before the next Claude release makes the answer "nothing."
Think I'm wrong? Tell me so. In public, so the wannabe experts can see it. I'll engage. That's how you find out who's actually built something and who's selling fairy dust.
Cheers,
Pete