like the moment when something clicks.
When you've been tweaking a recipe and finally nail the flavor. When you're debugging and suddenly see the root cause. When you're reading a paper and a concept from a completely different field connects. The domain changes, but that "this is it" feeling is the same.
I wanted to build a note app that supports the trial and error behind those moments. That's why I started building Graphium.
The limits of lab notebooks
When I was doing materials science research, I kept running into the same problem: experimental processes were hard to record and even harder to pass on. In grad school, different senior researchers taught me different preprocessing procedures for the same sample material. The know-how lived in people's heads, not in any shared system. We wrote things down in paper lab notebooks, but describing a process precisely is tedious, and searching through those notes later is nearly impossible. On top of that, the data that makes it into publications is just the tip of the iceberg — the trial-and-error data sitting in each lab usually goes unused and eventually disappears.
In 2019, I had a vague idea: "If we could record experimental processes in a structured way, maybe AI could make use of them." (I wrote about this problem in a blog post and a FIT2020 presentation at the time — both in Japanese.) That was the starting point for Graphium. But I had no concrete vision of what it would look like.
Pieces that didn't fit yet
Over the next few years, the pieces I needed started coming together in my hands, one by one.
In 2023, around the time GPT-4 came out, I discovered Niklas Luhmann's Zettelkasten — a method where you write individual notes in your own words, link them to each other, and let a network of thought grow over time. I was drawn to the idea that connecting notes could lead to new insights. At the same time, I felt the weight of "permanent notes" — the process of abstracting and restructuring raw notes into refined knowledge. "Could AI handle this part?" I thought. But I couldn't see how this connected to my 2019 problem yet.
In 2024, I came across BlockNote.js, a block-based editor framework. Each element — text, images, data — exists as an independent block with its own ID. I sensed potential in its extensibility. At that point, it felt like "an interesting piece of technology" to me personally, but I didn't yet have a concrete sense of how it would tie into my 2019 problem.
In 2025, I learned about PROV-DM, the W3C standard for describing provenance. It models relationships between Entities (things), Activities (actions), and Agents (actors) in a structured way. That year, a colleague and I wrote a paper called MatPROV, applying PROV-DM to structure the provenance of materials synthesis. It was accepted at a NeurIPS 2025 workshop. For the first time, the vague 2019 idea — "structured recording of experimental processes" — had a formal framework to attach to.
The pieces were accumulating. But I still couldn't see how to put them together into one thing.
The moment everything connected
In January 2026, a colleague suggested: "What if you could attach context labels to blocks?" A simple idea — give semantic meaning to individual blocks via labels. But the moment I heard it, the scattered pieces clicked into a single picture. Blocks have IDs. IDs mean you can attach labels. Labels mean you can auto-generate a PROV-DM provenance graph. Links between blocks become a Zettelkasten network. And if AI layers a knowledge base on top of that network, you can break through the permanent-note bottleneck.
It also helped that this was a time when AI could vibe-code. Once the ideas connected, I could immediately start building. I went from incubating a concept to implementing it almost overnight.
Then in April 2026, Andrej Karpathy proposed a design pattern called LLM Wiki — an approach where LLMs continuously build and update a Markdown wiki. As he put it: "LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass." The idea I'd been carrying since 2023 — "let AI handle Zettelkasten's permanent notes" — suddenly had a concrete implementation pattern.
Three layers of Graphium
These pieces combined into a structure where Graphium supports discovery through three layers.
Organize your thinking. Type @ to reference another note, and a network between your notes starts growing. Zettelkasten's "connect your thoughts through links" philosophy, built directly into the editor.
Accelerate discovery. As you accumulate daily notes, AI reads across them and auto-generates a knowledge layer draft. For example, scattered experimental findings across multiple notes get organized into a single synthesized page, which you can then review and edit. With minimal effort, a knowledge base equivalent to Zettelkasten's permanent notes grows over time.
Protect your discoveries. Attach labels like #Input or #Output to blocks, and a provenance graph is auto-generated, showing what went in, what steps were taken, and what came out. Back in grad school, different senior researchers taught me different preprocessing procedures and I couldn't reliably reproduce any of them — this is exactly the kind of tool I wished I'd had back then. Provenance is described using the W3C standard PROV-DM, and when you quote or derive new notes from existing blocks, those connections are structurally recorded too.
Without any action, it stays a simple note app
A key design principle: these features only appear when you need them.
Without labels, Graphium is a simple note app. Start using @ references and the link network becomes visible. Add # labels and provenance graphs appear. Enable AI and the knowledge layer starts growing. Complexity scales with your actions, step by step.
Structure reveals itself gradually, only to the extent you need it — a progressive design I want to preserve.
Just getting started
The first commit was on March 23, 2026 — just about a month ago. There's plenty left to build, but ideas I'd been carrying in fragments since 2019 are coming together into a single product, and that process itself has been a chain of discoveries.
Right now Graphium is built for individual discovery. But I have a feeling this structure could eventually extend toward formalizing tacit knowledge within teams and growing collective intelligence.
In this series, I'll walk through Graphium's design decisions one by one. Why I built it this way, what I chose not to build, what I'm still figuring out. I want to share the development process as it actually is.
Top comments (0)