During the months of building infrastructure for personal AI - the memory systems, the conversation persistence, the architectural insights documented in my previous posts - one aspect seemed like a straightforward task: feed my daemon my context so it could understand me.
I had backups for blog posts, I scraped decades of forum discussions to extract my posts, I had (some) work documents, presentations, emails, I had a good chunk of my library as ebooks on my hard drive, etc. The equation felt simple: process this content through an AI, and it would understand who I am.
But what does “understanding who I am” actually mean?
My first discovery was that “my context” wasn’t one thing but three fundamentally different categories:
The Public Bucket: these would be the books I had read, the movies I had watched, my courses from university or from various seminars I attnded during my career, papers that influenced me. This is content created by others, publicly available. The same copy of “Atlas Shrugged” that shaped my thinking exists on millions of hard drives.
But here’s the thing - my daemon doesn’t need to know what Ayn Rand wrote. It needs to know how those ideas connected to my existing thoughts, which chapters resonated, what arguments I rejected, how it changed my approach to recursive problems. The book is public; my relationship with it is personal.
The Personal Bucket: this is all the content I created: blog posts, code, presentations, emails, forum posts, you name it. This seems obviously necessary - it is literally my voice, my thoughts, my expression.
But I discovered raw content isn’t understanding. My angry email from 2015 doesn’t explain why I was angry. My code doesn’t reveal why I chose that architecture over another. My forum post assumes context I never made explicit - the failed project that taught me those lessons, the mentor who shaped that perspective, the specific pain I was solving for. And if I want full context, where does I stop? My Whatsapp account? My thanks on a birthday post on Facebook? My dating app’s message history?
The In-Between Bucket: This was another surprise question: what about the context in which my content sits. The forum threads where I debated for pages. The email chains where ideas developed through exchange. The meeting notes from projects where my contributions mixed with others'.
This content isn’t mine alone, but it shaped me. And I shaped it back. When someone on a forum challenged my position on photography gear and I spent three days crafting a response, their pushback became part of my thinking. The arguments that changed my mind, the collaborative problem-solving, the intellectual sparring - these interactions are an integral part of my intellectual development.
Each bucket raised different questions:
There is something else about in the personal AI vision: we are not children getting daemons at birth. We are adults trying to create instant intimacy with decades of history.
In Pullman’s world, daemons emerge alongside their humans, growing together, developing shared understanding through shared experience. Every formative moment is experienced together. The daemon doesn’t need to be told about your childhood trauma or your first love or your career pivots - it was there.
I am trying to create a daemon that suddenly needs to understand fifty years of intellectual development, career evolution, relationship patterns, accumulated wisdom and persistent blind spots - all at once.
This isn’t just a data ingestion problem. The context that makes me “me” wasn’t formed in isolation but through interaction, over time, in response to specific circumstances that no longer exist. My approach to system design comes from battles fought in companies that no longer exist, with technologies now obsolete, solving problems that modern stacks handle automatically. But those battles still shape how I think.
How do you teach that to an AI? Not the facts of what happened, but how those experiences layer into current thinking?
I spent weeks building content processing pipelines. Semantic extraction. Knowledge graphs. Vector embeddings. I could feed thousands of documents through sophisticated AI models to extract “key insights” and “core concepts.”
But I kept hitting the same wall: processing isn’t understanding. Claude could tell that I’m a passionate skier but not how it shaped my approach to risk. Claude could tell what metaphors I keep revolving around, but not where my habit comes from. The AI could catalog my thoughts but couldn’t understand my thinking.
I even tried another approach: feed it with biography. Start with a CV, lay out the narrative of my life, key events, turning points. Give the AI the story that connects all the content.
But whose story would I tell? The one where I am the hero of my own journey? The one where I am constantly learning from mistakes? The one where random chance shaped more than I would like to admit? We all carry multiple autobiographies, deployed strategically depending on context and audience.
Even if I could write an “objective” biography, it would be a lie through omission. The most formative experiences often can’t be shared - the betrayal by a trusted friend, the project that failed due to politics, the stupid stuff done as a teenager that I never told my parents. These gaps in the narrative aren’t just missing data; they are key anchors that shape everything around them.
The deeper I went, the more I realized I was trying to solve the wrong problem. I assume personal AI needs explicit knowledge about me - my documents, my history, my stated preferences. But what makes someone know you isn’t the facts they can recite but the patterns they recognize.
My partner doesn’t know me because she has memorized my biography. She knows me because she recognizes when I am procrastinating versus when I am processing. She can tell the difference between my “I’m fine” that means I am fine and my “I’m fine” that means I am drowning. That knowledge was not downloaded; it was developed through thousands of micro-interactions.
But that kind of knowledge seems to require exactly what current AI architectures can’t do: learning continuously from interaction rather than processing static content.
This brings me to an uncomfortable question: what kind of “knowing” am I even trying to achieve?
There is knowing-about: facts, preferences, history. “He knows SAP, grew up in France, has three kids.” This is what we get from processing content.
There is knowing-how: patterns, tendencies, responses. “He overthinks technical decisions when he is unsure about requirements.” This emerges from observation over time.
There is knowing-with: shared context, mutual understanding, collaborative thinking. “We have worked through this kind of problem before.” This only comes from actual interaction.
My approach optimized for knowing-about, and I thought I could feed it enough content to help with knowing-how when what I expect for my dsaemon is knowing-with. But I can’t be bootstrap it from content. It has to be grown through relationship.
So I am back to a version of my original question, but inverted. Instead of “what should my daemon know before it meets me?” I am wondering “what kind of knowing can I actually bootstrap, and what has to be developed?”
Maybe I’ve been thinking about this backwards. Instead of trying to upload “myself” into an AI’s understanding, maybe personal AI needs to start nearly blank and learn me through interaction - like any other relationship.
But if that is true, the fifty years of context I wanted to feed it is not training data. It is conversation material. Not “here’s who I am,” but “let me tell you about this experience.” Not processing my blog posts to extract my voice, but reading them together and discussing what I was thinking when I wrote them.
This radically changes what personal AI infrastructure needs to optimize for. Not bulk content processing but continuous learning. Not knowledge extraction but relationship development. Not understanding from data but understanding through interaction.
Which brings me back to the architectural problem: current AI systems are built for the former, not the latter. We are trying to speed-run relationships using tools designed for stateless services.
The question is not “what does it mean for an AI to know me?”
The question is “what does it mean for an AI to be capable of getting to know me?”
And that is a fundamentally different problem than the one we are trying to solve.