Expert Take — Hans Schepers
AI is the smartest assistant I’ve ever had — and also the most unreliable.
Last updated: April 2026
Where It Went Wrong
In WordPress, which I use to build my website, multiple user accounts had somehow been created for me — hans-editor and hans. No idea why, but there they were. It seemed logical to remove that first oddly named account, so after “consulting” with Claude I went ahead and did it. To my surprise, Claude advised me an hour later to start using the hans-editor account again. I was baffled. Was I getting confused?
No, I was right. Claude had stored in its history that hans-editor could be deleted, but had also stored, a few days earlier, that the account had to be created. Both memories coexisted. Only after I instructed Claude to archive all obsolete memories did it start remembering only the correct things.
I had assumed that “a project” in Claude automatically meant continuity. That turned out not to be how it works.
This was the beginning of a series of discoveries about how AI tools actually function in a real project.
What Kept Going Wrong Structurally
Important decisions are forgotten over and over again. I created a voice profile as a skill in Claude. With it I can, for example, record something and have Claude rewrite it while accounting for my specific tone of voice. But after generating such a text and starting a new chat session in the same project, the automatic use of my voice profile was simply gone. Every new chat turned out to be a new instance, and I had to ask for it every time. There was no automatic bridge to what I had already explained in a previous chat — not even within one Cowork project.
Valuable data gets lost during chat compressions. The longer a chat session, the sloppier Claude becomes. With long sessions, Claude automatically compresses the conversation history. It took days before I realized that valuable information was disappearing during that compression. Information about errors we had explicitly evaluated together. Or information about why certain decisions had been made. These details vanished without warning. So I had to figure out how to save information before that automatic compression kicked in.
Assumptions are often presented as facts. Many times, Claude wrote pure nonsense with a tone so self assured – I didn’t dare to doubt. If you’re not paying close attention to every detail, those hallucinated “facts” get treated as truth. I made the mistake multiple times of not being sure exactly what had been discussed and giving Claude the benefit of the doubt. Wrong. Don’t do it. Instruct Claude to archive all your history and search it when in doubt.
Why This Is Particularly Problematic for Project Work
Asking standalone questions to an LLM works fine. That’s what it’s built for. But a project is something different. A project has continuity, decisions that build on each other, and corrections that need to persist.
Every correction I had to repeat cost me time and undermined the entire idea of collaboration. If the AI starts fresh after every conversation, a project is just a series of standalone chats in a folder. That’s useful, but it’s not what I expected when I started “collaborating with AI on a project.”
It confirms something I’d suspected for a while: AI knows everything but understands nothing. Claude had all the information I’d given it. But without a system to learn from it, that information just evaporated.
What I Had to Build
Here’s what’s in place now:
A permanent memory system where every decision — or correction to one — is immediately saved to an indexed file. Not at the end of a session, but the moment the decision is made. That memory gets automatically loaded at the start of every new session.
A session protocol with three mandatory phases: session start (read the project status, confirm all rules are loaded), during the session (save corrections in real time, update every hour), and session end (archive what was done, update a handover file that shows exactly where to pick up in the next session). If I’d done this from the beginning, it would have saved me a lot of time — and a lot of frustration.
A start confirmation where Claude confirms at the beginning of each session that it has read all the context and clearly summarizes where you are in the project, including the next steps.
Is this a lot of work? Yes. Is it worth it? For this project, yes — I notice the difference. But I also understand why most people don’t do this. It requires a different mental model of what AI tools actually are.
The Mental Model I’ve Updated
I now think of Claude as a brilliant colleague who shows up at the office every morning with amnesia. Intelligent, capable, fast — but without any memories of yesterday. If you don’t build a system that constantly feeds and refreshes those memories, the collaboration just doesn’t work. Not for me, at least.
That’s not a criticism of Claude. It’s how LLMs work. But it’s also something you have to discover for yourself. My expectation was that it “just works,” and when it doesn’t work as expected, I would usually think I was the dumb one in the room.
What This Means for Your AI Projects
If you’re considering using AI structurally for an ongoing project — content, strategy, research, whatever — these are the questions I’d ask myself first:
How do I save corrections so they’re available in the next session? How do I prevent information from being lost in long chats when compression kicks in? Which information needs to always be available, regardless of how long the session runs? And: which tasks am I imagining AI handles autonomously, while in practice they require human guidance?
I hope this gives you something useful. And I’d genuinely like to know: what false assumptions about AI intelligence ended up creating extra work for you?
Sources: Personal project experience GEO Masterplan (March–April 2026). Background on LLM context windows and compaction: Anthropic Claude documentation. Research on AI assumptions and hallucinations: GPT-4 Technical Report, OpenAI (2023).
FAQ
Does Claude not have memory at all?
Not by default. Every new chat starts with a blank slate, even a project in Claude Cowork. There are ways to add persistent memory — like the auto-memory system I use — but I had to set those up for myself.
What exactly is context compression?
In long conversations, an LLM has a maximum amount of text it can actively process (the context window). When a conversation gets too long, the model automatically compresses the older parts. Details, nuances, and corrections can get lost in that process — without any notification.
Is this problem specific to Claude, or does it apply to all AI tools?
It applies to varying degrees to all LLM-based tools. The specific mechanisms differ per platform, but the fundamental issue — no persistent memory between sessions — is widespread. GPT, Gemini, and other models have the same structural limitation.
How do I notice that context compression has happened?
I often only noticed it when the output contained something that contradicted something that had been discussed earlier. My advice: build checkpoints into your sessions where you verify that the AI still has all the relevant context. And: keep sessions as short as possible.
Read also: Everything a B2B Marketer Should Know About GEO in 2026 · What is GEO? · What is AI Search?