After months of building infrastructure and hitting walls, after documenting why Personal AI is harder than it looks, after questioning what it even means for an AI to “know” someone, I feel I am stuck. I feel like I am at a crossroads with multiple paths forward, none clearly superior, each with its own trade-offs that will only become apparent after you have traveled far into it.
This paper is an inflection point in my series. My previous paper have been written after the fact, they documented what I discovered. Now we have caught up - this paper documents the decision point itself - the messy middle where theoretical understanding meets practical choices, where vision meets implementation constraints, where perfect becomes the enemy of good enough.
Let me start with where I am at, what actually exists.
I have a functional multi-model chat system running on n8n. It is not pretty - the UI would make any product designer’s eyes bleed - but it works. I can have a conversation with Claude on Monday, continue it with GPT on Wednesday, reference both while talking to my local Llama model on Friday. Every conversation gets stored in a database with a consistent schema. The persistence is real, even if retrieval requires fumbling around.
This might not sound like much, but it solves one fundamental problem: conversation fragmentation. My discussions don’t vanish into the ether when I close a browser tab. They accumulate, available for reference, ready to build upon. It is the difference between constantly starting over and actually building something. Yes, all major AI bots provide similar feature nowadays, but they are siloed.
I’ve scraped and stored twenty years of forum posts across three platforms. These aren’t just any posts - they are my authentic voice across decades, me arguing, me engaging in various communities. Thousands of posts, stored, searchable, ready for… something.
My blog posts sit in markdown files with proper metadata. Easy to parse, ready to process.
Five hundred books that “mattered to me” converted to text files. Some I have read multiple times, all that influenced specific periods of my life, all chosen because they left marks on my thinking. Megabytes of potential knowledge waiting to be understood by something that understands me.
Ten years of work documents - presentations I created, documents I developed, policies I have implemented, training documents I had to learn, minutes of meetings I have been part of, etc. Not everything of course, but a subset that has influenced me and shaped who I am professionally.
But here’s the thing: all this content just sits there. My chat system doesn’t know about my forum posts. It can’t reference my blog writing style. It hasn’t read the books that shaped me or learned from the documents I have produced.
I have infrastructure and I have content, but they are like two people at a party who haven’t been introduced. I have advocated in the previous papers about the potential for meaningful interaction that it enables, but the connection hasn’t been made.
In this paper , I explored what it means for an AI to “know” me. Now I face the practical version of that philosophical question: how do I bridge the gap between raw content and useful augmentation?
The naive approach seemed obvious: process everything through AI, extract insights, build a knowledge base, connect it to chat. But every attempt reveals the same catch-22: to extract personally meaningful insights from my content, the system needs to already understand me. But it can’t understand me without processing the content that shaped me.
Take my forum posts. An AI could process them to extract topics I discuss, positions I take, how my views evolved. But what it can’t extract is why my approach to photography evolved from being a gear-head to favoring composition. The context that makes my content meaningful isn’t in the content itself.
Standing here with working infrastructure and unprocessed content, I see three fundamentally different approaches. Each has merit. None is obviously correct.
Forget bulk content processing. Focus on making the chat system genuinely useful as a tool I want to use daily. Better conversation retrieval, automatic summarization of past discussions, smart context management that knows which previous conversations might be relevant.
The appeal: This builds the daemon through interaction rather than preprocessing. Every conversation teaches it something about me. The relationship develops naturally, like any relationship does - through engagement over time, not through downloading my history.
The hesitation: This feels like replicating existing products (Claude’s Projects, ChatGPT’s Memory) with inferior tools. Yes, I would own the data and control the system, but am I just building a worse version of what already exists? My time is limited - should I spend it creating basic features that others have already perfected, just for the sake of data sovereignty?
There is something else: this path accepts the current paradigm of base model plus context. I would be building better infrastructure around fundamental limitations rather than addressing the limitations themselves.
But that also would mean a working product, a feel of success, likely stenghtening my knowledge of the field, exploring new tools to gain time (stuff such as librechat.ai look promising), so all in all, progress.
Pick one specific area - probably my forum posts since they are already structured - and go deep. Try different processing approaches: fine-tuning, RAG, knowledge extraction, pattern analysis. Not to solve everything, but to learn what actually helps.
The appeal: This is manageable, measurable, and might reveal insights about the larger problem. If I can make my daemon understand my technical discussions well enough to genuinely help with similar problems, I will have learned something valuable that could extend to other domains.
The hesitation: I might lack the skills. Fine-tuning models, building effective RAG systems, creating meaningful embeddings - these require expertise I am still developing and I can see the rabbit hole. More fundamentally, I haven’t properly defined what success looks like. What does it mean for an AI to successfully “understand” my forum posts? How would I even measure that?
I have started to learn more about knowledge graphs. The premise is compelling: instead of treating content as isolated documents, map the relationships between concepts, creating a semantic network that AI can navigate. Yes, these are solutions developed for corporate knowledge management, but is “personal knowledge management” the missing link for my daemon?
For my content, this might mean identifying entities (people, projects, technologies), relationships (worked on, influenced by, responded to), and temporal connections (evolved into, replaced by, learned from). The AI wouldn’t just retrieve relevant documents but understand how ideas connect across my intellectual history.
The appeal: This addresses something RAG misses - the connections between disparate pieces of content. That avalanche safety course that influences my approach to risk management practise? A knowledge graph could capture that non-obvious relationship.
The hesitation: Knowledge graphs might be solving the wrong problem entirely. They are excellent for mapping explicit relationships between entities, but my daemon needs to understand implicit patterns in how I think. The insight that “he over-engineers when uncertain about requirements” isn’t a graph relationship - it is a pattern recognized across hundreds of interactions.
There is also the effort question. Building a meaningful knowledge graph from my content would require not just technical implementation but extensive manual curation. Is the benefit worth the investment?
As I consider these paths, a deeper question emerges: am I trying to build the right thing the wrong way, or the wrong thing entirely?
Current AI architectures - base models that are trained once then frozen, context windows that temporarily hold information, RAG systems that retrieve but don’t learn - might be fundamentally unsuited for personal AI. I am trying to create continuous relationship using tools designed for stateless interactions.
Yet these are the tools available. The perfect architecture for personal AI - models that truly learn from interaction, that update their parameters based on conversations, that develop genuine understanding over time - doesn’t exist yet. Or rather, it exists in research papers but not in anything I can actually run on my hardware.
So the question becomes: do I work within current limitations, building the best personal AI possible with frozen models and external memory? Or do I wait for (or contribute to) architectural advances that could enable true cognitive companionship?
I’m realizing that “minimum viable soul” might mean something different than I originally thought. Not a simplified version of the full daemon vision, but a system that provides genuine value within current constraints while building toward something greater.
Success might be:
Or success might be discovering that current approaches can’t bridge the gap between tool and companion, that we need fundamental advances before personal AI becomes possible. That would be valuable knowledge too.
I could end this paper with a decisive choice, a clear path forward. But that would be lying about my state of mind. The truth is that standing at this crossroads, I don’t know which path leads toward genuine personal AI and which leads to elaborate dead ends. What I do know is that the exploration itself has value. Each experiment, even failed ones, will allow me to learn valuable knowledge, but the cost of experimentation - mostly my own time - is not to be overlooked.
So instead of choosing a path, I guess I’ll keep exploring along each of these paths simultaneously, documenting not just the destinations but the journey itself. The minimum viable soul might not be a single system but a collection of experiments that collectively point toward what is possible and what is still missing.
If you’re reading this, whether because you’re building your own personal AI or because you are interesed in the topic, I have questions that might be worth exploring together:
The minimum viable soul will be a journey of progressive approximation, each experiment bringing us closer to understanding what we are really trying to build.