My Best AI Decision

I have a problem that some know well; The gap between what you know and what you can actually find when you need it. I have spent years building what the productivity crowd calls a "second brain." Obsidian vaults, Apple Notes graveyards, Brave reading lists, notebooks, collections of bookmarks I swore I would organize later. Every single one of them followed the same pattern. Enthusiasm at the start, a burst of tidy structure, and then... the slow decay. Notes pile up in random places. Links break. The table of contents becomes fiction. Within a few months, I was back to using a search engine and my own unreliable memory instead of the expansive system I had built. I am not good at keeping things tidy. I will say that plainly. I have tried. I have bought the templates, watched the YouTube tutorials, done the Sunday evening organization sessions. The notes got filed, but they never got found again. Something was missing, and for the longest time I assumed the problem was me. It turns out the problem was that I was trying to do this alone. The Insight That Changed Everything Every system I have ever used assumes that a human being will come back later to prune the dead branches, fix the broken links, and surface the forgotten ideas. But nobody does that. Nobody has the time, and even if they did, it is tedious work. It is the digital equivalent of weeding a garden... and I am the guy who lets the tomatoes die because I forgot to water them. The insight hit me gradually. The power was not in creating the wiki. It was in having a system that grows and tends it automatically. A place where dropping in a thought is the beginning of a process, not the end of one. Where notes do not go to be stored... they go to be built upon. The architecture is simple, which is probably why it works. I use Obsidian as the vault, and everything inside it is plain Markdown. No proprietary formats, no databases I cannot read with a text editor, no vendor lock-in. Markdown has become the language of AI, and that matters more than I initially realized. An AI agent can read it, write it, link it, and reason about it without any special tools. The vault is organized into sections that map to how I actually think. A wiki for concepts and entities. An inbox for raw thoughts I have not processed yet. A research folder for deep dives. Logs for tracking what changed and when. It is an operating model that grew out of what I actually needed. But the real structure is not the folders. It is the three layers running underneath them. There is the human layer, which is just me dropping things in. A question I had before bed. A link from a conversation. A half-formed thought I did not want to lose. While most of the images you'll see below are screenshots of Hermes' CLI, these inbox items can be dropped directly into Obsidian through the app on my phone, via Telegram, or a web-ui. There is zero friction for getting an item actioned. It's time to introduce you to Percival. Yes, I admit I have a problem anthropomorphizing my agents... it just helps me to have a mental image of who I'm talking to! This is the AI agent layer, which is running on the Hermes agent harness. My archivist, researcher, and occasional critic... his SOUL.md is written in the style of a haughty, old-man British guy who knows everything and has zero problem letting you know that. Percival knows the vault better than I do because he built half of it. He knows my preferences, my projects, my blind spots. He's the one autonomous AI agent I have that knows everything and can see everything (but has zero write access outside of the vault). Every note, every scribble, every screenshot and random note from a video game, every recipe and every smoked pork belly journal, Percival has seen, read, and likely fixed somehow. And there is the automation layer, the cron jobs and scheduled tasks that make sure the system never depends on me remembering to check the inbox. Daily, Without Exception Every morning at nine and every evening at nine, a job runs that scans the inbox for anything new. It does not wait for me to notice. It does not require a click or a command. It just runs. What it does depends on what it finds. A two-line note about a topic I am curious about might get expanded into a structured wiki entry with cross-references to related concepts. A link I dropped in might trigger a summary and a decision about where it belongs. A question might get queued for deep research. A stray thought might get filed as a concept with connections to half a dozen other ideas I had forgotten I wrote down. The inbox is the front door, but it is not a queue. It is the starting point for everything else. What goes in as a raw, unformed note comes out as something the rest of the system can use. Structured, linked, filed, and findable. My favorite automation is one I've only begun experimenting with recently. It's very broad, very open ended. It's a simple "surprise" task that asks Percival to wake every morning at 6:30am, take everything he knows about me and what I am interested in, and goes out and compiles a report on something he thinks I might find interesting. It's open-ended for a reason, as I believe a little whimsy and randomness are important parts of life. AI can be so much more than just summarizing a meeting. There are additional automations (cron jobs) for Percival. Things like a de-duplication layer that goes through the vault and looks for similar entries and combines them when it makes sense. Tag-cleaning, link management, a graduating process from idea to long-term wiki entry. All of this is automated now. I used to lose ideas because I never got around to processing them. Now I do not have to be the one who processes them. When a Question Deserves More Than an Answer Sometimes I drop a question in the inbox that is too big for a single wiki entry. A topic I need to understand deeply, not just record. For those, the system runs deep research. It is an eight-phase pipeline that starts with twelve concurrent search angles, pulling from sources across the web. It registers every source it finds. It writes in sections, progressively, building arguments rather than dumping information. It tracks citations so I can verify anything that sounds surprising. And I don't do a single bit of it. The result is not a collection of raw notes I still have to sort through. It is a finished report, structured into sections with cross-references to my existing wiki knowledge. It builds on what I already know instead of pretending I am starting from zero. It angles the information to me; if I leave a note about an interesting new piece of open source software, it automatically links to and leverages existing vault information about my homelab into the report. I have had 25,000 word reports waiting for me at eight in the morning that started as a two-line question I scribbled down at eleven the night before. That still feels a little like magic, even though I know exactly how it works. The Nervous System None of this would work without the linking layer. The concept comes from Andrej Karpathy's llm-wiki, which he built to solve a similar problem at a much larger scale. The idea is straightforward but transformative. Every concept in the vault should be linked to every other concept that relates to it. The system enforces this. It maintains the graph. It catches orphan pages that have no incoming links. It keeps the knowledge base navigable not just by search, but by following connections from one idea to the next. The difference between a folder of Markdown files and a real knowledge base is exactly this. Without links, you have a filing cabinet. With them, you have a map. Karpathy's insight was that this linking structure is what makes a wiki genuinely useful for both humans and AI. I can follow a thread of thought across dozens of pages, and so can Percival. When he writes a new report, he knows what I already believe because he can see how it connects to everything else I have written. The context is never missing because the context is woven into the structure itself. Here is the part I find most remarkable, and I did not plan for it: The system gets smarter the more I use it. Every report that gets written becomes part of the wiki. Every new concept gets linked to existing ones. Every question I ask gets answered in a way that references what I already know. So the next time I ask a question, the answer is better targeted because the system understands my context more deeply. The wiki is not just an output. It is an input to every future piece of research. The inbox feeds the research. The research feeds the wiki. The wiki feeds better research. It is a loop, and it compounds. Percival knows my setup, my projects, my recurring questions. He knows which sources I trust and which ones I do not. He knows when I am asking about something for work versus something I am just curious about. That context does not come from a prompt. It comes from the fact that he has been tending this garden with me for months. What This Actually Replaces When people talk about AI use cases, they usually mention coding assistants, or summarizing meetings (all I heard for MONTHS at work was about summarizing meetings!), or generating first drafts of emails. Those are fine. I use some of them myself. But the number one use case of AI in my life, by a very wide margin, is this. Having a researcher who understands my context, my vault, my situation... and who turns my scattered curiosity into something I can actually use. I am not a particularly organized person. I have already admitted that. What I am is deeply curious, I love learning and I have a lot of questions. Questions about AI policy and privacy, about how large language models actually work, about why LeBron James is objectively the GOAT, about the history of ideas I only half understand. Before this system, those questions would rattle around in my head until I got distracted by something else. Now they get answered. Thoroughly. With connections to things I wrote six months ago that I had completely forgotten about. That is the use case that hit different. Not the novelty of watching AI write something. The usefulness of having that writing be genuinely relevant to me. I should be clear that I did not invent any of this from scratch. As I already mentioned the conceptual foundation for the linking layer comes from Andrej Karpathy (co-founder of OpenAI, by the way). I adapted it to my own vault, but the core idea is his. The deep research pipeline is based on a skill I found on GitHub, which provided the citation manager, the eight-phase structure, and the progressive writing approach. I have modified it heavily to fit my workflow and my voice, but the original framework deserves the credit. Obsidian provides the vault architecture that makes all of this possible. Markdown provides the universal language that both I and my AI agent can read and write without translation layers. And the agent infrastructure itself runs on the Hermes Agent framework, which handles the scheduling, tool access, and persistent memory that let Percival actually know who I am from one day to the next. This is still the beginning. The system works well enough that I rely on it daily, but there is more I want to do. Very soon (days) I plan on implementing qmd, which essentially turns the vault into an even easier to search database via CLI. I am exploring additional agents for specific tasks. A specialist for technical documentation. Another for tracking long-term research threads that span multiple reports. I want better publishing workflows, so that finished research can move from the vault to a blog post or a presentation without manual copying and reformatting. I see a future where Percival "employs" an army of sub-agent researchers and workers that have a narrower focus, which ultimately creates better results. Mostly, though, I want to keep refining the feedback loop. Making the research better targeted. Making the wiki more navigable. Making the whole system feel less like a tool I use and more like a natural extension of how I think. Perhaps you're like me and don't do so well with digital organization. Perhaps you're still not sure about AI and haven't really found a use-case yet. Maybe you're a CEO, CFO, COO, CTO and are just looking for some practical assistance gathering information on things that maybe you don't want to even ask an assistant about (I get it, we've all wanted to know the best way to make money off turnips in Animal Crossing before... your secret is safe with me). This is a fairly safe way to dip your toes in the water of actual life-improving, semi-autonomous agentic AI. I use Hermes because I find it to be incredibly powerful and customizable. But most platforms have some sort of task layer baked in now. With some clever prompting and scheduling, all of this should be doable in just about any tool or platform. Skills make it easier than ever to set things like this up. Embrace your curiosity and keep learning... even if it is just about turnip exchange rates in an old Nintendo Switch game. --- References: - Obsidian - Andrej Karpathy's llm-wiki - Deep Research by 199-biotechnologies - Nous Research Hermes Agent - qmd