I am not an AI researcher. I am a systems architect who tinkers with technology, and I have spent months trying to build something that shouldn’t be this hard: an AI that actually knows me. Not in the creepy surveillance way, but as a genuine cognitive companion that learns and grows through our interactions.
What I discovered was that I kept hitting the same walls regardless of approach - I may be bad at implementation, but that doesn’t account for everything. These walls felt fundamental, not incidental. Eventually I realized I wasn’t failing at building personal AI - I was discovering that current architectures might not support what I was trying to build.
This is one tinkerer’s perspective on why that might be.
My goal seemed straightforward: create an AI system that maintains continuity across conversations, learns from our interactions, and processes my personal content to deeply understand my context, my cultural background, what shaped me. I had working pieces and the technical integration worked fine [link to the Journey paper].
But something fundamental was missing. Despite perfect memory of our conversations, the AI never really learned anything about me. Each interaction still required extensive context-setting. I could automate part of this context-setting, replicating features similar to Claude’s projects, this made it easier ny creating some sort of memory, but the system still had no understanding of me.
Why? Because the model itself never changes after training. Current LLMs follow this pattern:
This works great and serves millions of users - you get consistent, “predictable” behavior. And adjusting the model to make it more specific is not news, it has been already addressed through multiple solutions:
All these solutions have their use: RAG is a great way to build a chatbot and enrich it with corporate knowledge, context windows are a necessity to coding or deep research agents, fine-tuning is how you can get specialized and truly competent models for use cases such as healthcare or law. But it seems fundamentally at odds with personal AI that should learn from every interaction and be truly tailored to an individual.
What I think I need is fundamentally different: a model that learns continuously from our interactions. Not just storing memories externally, but actually updating its parameters based on our conversations. The model I talk to after a year should be structurally different from the one I started with. I mean, I expect Youtube and others to suggest stuff based on my interests, I want my AI to do no less. Hence the importance of cognitive sovereignty I mentioned in my first paper [link to the Manifesto].
Here’s what struck me: my personal AI doesn’t need to know everything about everything. Yes, there are some basic elements required - understand the language(s) I am speaking, creating and consuming content in, feels like a prerequisite. But mostly, it needs to know about me, my interests, my patterns. That’s a drastically smaller surface area than “all human knowledge”.
And this makes the scale question completely different: it doesn’t require a big cluster of industrial-grade of GPUs for some time to be trained, but it needs continuous access to some sort of computing for background processing. And everyone of us has access to lots of cheap and idle compute. This again is far from new: remember the SETI screen savers back in the day? A modern smartphone has tons of computing power, a gaming PC provides significant (and mostly idle) compute capability. Not enough to train a GPT-5 model, not even close, but maybe enough to continuously adapt a smaller model to one person? The economics might work if we’re not trying to create universal knowledge but personal understanding.
The ML community calls this “continuous learning” and apparently it’s hard. There is something called “catastrophic forgetting” where learning new things makes models forget old things. But I wonder if this is a problem specifically because current architectures assume learning happens once then stops. Maybe architectures designed for continuous learning from the start would handle this differently?
Again, I’m not suggesting current approaches are wrong - I myself am a happy user of Claude and Gemini. A model serving millions needs to be frozen for consistency. Cloud providers need predictable behavior. These are valid requirements for those use cases.
But personal AI has different requirements that might need different architecture. Just like personal computers required different architecture than mainframes, not just smaller mainframes.
The question I keep returning to: are we trying to build personal relationships with architectures designed for stateless services?
If my experience is indicative of a broader limitation, it suggests personal AI might require:
Continuous Evolution: Models that change through use, not just through training. Your AI after five years of interaction would be fundamentally different from where it started. But there would be a full continuity from then to now.
Personal Scale: Instead of massive models knowing everything, smaller models that deeply understand individuals. Quality of understanding over quantity of knowledge.
Different Economics: Computing distributed across time rather than concentrated in training bursts. Hardware individuals already own rather than data centers.
Unified Memory: Understanding encoded in the model’s structure, not just retrieved from external stores.
Easier said than done, I am totally aware of it, and even beyond, critical questions remain, just to name a few:
These feel like engineering challenges rather than fundamental impossibilities - to me. I could be wrong. Perhaps I’ve misunderstood the technology or missed obvious solutions. But if others are hitting similar walls, maybe it’s worth asking whether we’re working around architectural limitations that shouldn’t exist. The capability for personal AI seems within reach. Current models can engage in sophisticated reasoning and understanding. What appears to be missing is architecture that allows that capability to become truly personal through continuous learning rather than frozen deployment.