Not all nodes are created equal
You have been building a knowledge graph — connecting ideas, linking notes, threading concepts across domains. By now you have hundreds of nodes. Maybe thousands. And if you look at the graph with fresh eyes, a pattern jumps out: a small number of nodes have far more connections than the rest. They sit at the center of dense clusters. Remove one and the graph fractures.
These are your hub nodes. They are not hubs because you decided they should be important. They are hubs because the structure of your knowledge made them so. Every new idea you added that touched "feedback loops" or "cognitive load" or "incentive structures" added another edge to those nodes. Over time, the rich got richer. The concepts that connect to many things attracted even more connections.
This is not a quirk of your particular graph. It is a law of network structure, and understanding it changes how you maintain your thinking infrastructure.
The science of hubs: why networks aren't flat
In 1999, Albert-Laszlo Barabasi and Reka Albert published a paper that reshaped network science. They mapped the World Wide Web and found something that contradicted the prevailing model: the distribution of links was not random. A tiny fraction of pages had thousands of incoming links while the vast majority had almost none. The web followed a power law — the same mathematical pattern Vilfredo Pareto observed in 1896 when he found that 20% of Italians owned 80% of the land.
Barabasi called these densely connected pages "hubs" and the networks they inhabit "scale-free networks." The mechanism that produces hubs is called preferential attachment: when a new node joins a network, it is more likely to connect to nodes that already have many connections. The rich get richer. A new Wikipedia editor is more likely to link to the article on "Evolution" (which has thousands of inbound links) than to an article on an obscure species of beetle (which has three). Each new link increases the probability of the next one.
This is not just a web phenomenon. Scale-free networks with hub structures appear in protein interaction networks, airline route maps, citation graphs, social networks, and neural connectivity patterns. The pattern is so universal that Barabasi titled his popular book simply Linked: the architecture of hubs and preferential attachment describes how complex systems organize themselves across every domain we have studied.
Your personal knowledge graph follows the same dynamics. You did not plan for "mental models" to become a hub node with forty connections. But every time you learned something new — about decision-making, about systems thinking, about cognitive bias — you naturally linked it to that concept. Preferential attachment happened organically, and the result is a power-law distribution: a handful of nodes carry a disproportionate share of your graph's connectivity.
How to find your hubs
There are two approaches: computational and intuitive.
Computationally, if your tool supports it (Obsidian's graph view, Roam's page references, Logseq's linked references), sort your notes by inbound link count. The nodes at the top of that list are your hubs. In graph theory, this measure is called "degree centrality" — the simplest centrality metric, counting how many edges connect to a node.
Google's PageRank algorithm — invented by Larry Page and Sergey Brin in 1998 — goes further. PageRank does not just count connections; it weights them by the importance of the connecting nodes. A link from a highly connected page counts more than a link from an orphan page. This recursive definition captures something degree centrality misses: a hub that connects to other hubs is more important than a hub that connects to peripheral nodes. Applied to your knowledge graph, this means the note on "feedback loops" that links to "systems thinking," "cognitive bias," and "habit formation" (all hubs themselves) is more structurally important than a note with the same number of links to minor, isolated ideas.
Intuitively, your hubs are the concepts you keep returning to. They are the ideas you reference in conversation, the frameworks you apply across domains, the notes you keep updating because new material keeps touching them. You probably already know what they are. The exercise is to verify your intuition against the data — and often you will find surprises. A concept you thought was peripheral may have quietly accumulated connections, or a concept you treat as central may have fewer links than you assumed.
Why hubs matter: the keystone species analogy
In 1963, the ecologist Robert Paine began an experiment on an 8-meter stretch of rocky shore in Makah Bay, Washington. He systematically removed a single predator — the sea star Pisaster ochraceus — from tidal pools while leaving adjacent control pools untouched.
The results were devastating. Within a year, the number of species in the removal zone had halved. After ten years, what had been a diverse ecosystem of fifteen species became a monoculture of mussels. The starfish was not the most abundant species. It was not the largest. But it was the most connected — the species whose interactions with other species structured the entire community. Remove it and the system collapsed.
Paine coined the term "keystone species" for this phenomenon, borrowing from architecture: the keystone is the single stone at the top of an arch that, once placed, holds the entire structure together. Remove it and the arch falls.
Your hub nodes are the keystone species of your knowledge graph. They are not valuable because they contain the most text or took the longest to write. They are valuable because they hold the structure together. Remove your note on "second-order effects" and every note that linked to it loses a connection, the paths between those notes lengthen, and clusters that were bridged through that hub become isolated islands.
This has a direct practical consequence: hub nodes deserve disproportionate maintenance. Most knowledge management advice treats all notes as interchangeable — review everything periodically, keep everything tidy. But the keystone species insight says otherwise. A poorly maintained peripheral note with two connections costs you almost nothing. A poorly maintained hub node with thirty connections degrades the integrity of your entire graph.
Attention sinks: the AI parallel
Your knowledge graph is not the only system where a few nodes attract disproportionate attention. Researchers studying large language models have discovered a strikingly similar phenomenon inside transformer architectures.
In 2023, Xiao et al. identified what they called "attention sinks" — specific tokens (usually the first token in a sequence or special delimiter tokens like [CLS]) that systematically attract a disproportionate share of attention weight across multiple attention heads and layers, regardless of their semantic content. These tokens function as structural anchors: the softmax normalization in attention layers forces the model to allocate attention somewhere, and these universal, always-visible tokens become the default recipients.
The parallel to hub nodes is instructive. In both your knowledge graph and an AI's attention mechanism, a small number of positions absorb most of the system's relational energy. The reason is the same in both cases: when a system must distribute limited resources (links, attention weight) across many items, structural dynamics concentrate those resources on a few high-connectivity points. The system does not choose its hubs through deliberation. Hubs emerge from the constraints of the architecture.
The practical lesson: just as LLM researchers must understand attention sinks to build efficient streaming systems, you must understand your hub nodes to build efficient knowledge systems.
Maps of Content: the intentional hub
Nick Milo, creator of the Linking Your Thinking framework, formalized a practice that directly operationalizes hub node theory. He calls them Maps of Content (MOCs) — notes that function as curated indexes for a specific topic, linking to all the relevant notes in that domain.
A MOC is an intentional hub. Instead of waiting for a concept to organically accumulate connections through preferential attachment, you create a note whose explicit purpose is to serve as a central navigation point. Milo describes MOCs as "workbenches" — spaces where you can see all the notes related to a topic laid out together, available for combination and synthesis.
This is valuable, but it comes with a risk. An organic hub earns its centrality because the concept it represents genuinely connects to many things. An artificial hub — an index page that links to everything in a category — can create the appearance of centrality without the underlying conceptual density. The note titled "Psychology" that links to every psychology-related note in your vault is an index, not a hub. It does not represent a concept that your other ideas genuinely depend on.
The strongest practice combines both approaches. Let organic hubs emerge through natural linking. Then periodically create MOCs for the domains where you need better navigability. But always ask: does this hub exist because the concept is genuinely central to my thinking, or just because I created a tidy index? The first is load-bearing infrastructure. The second is a filing cabinet.
The power law in your graph
If your knowledge graph has meaningful structure, it will follow a power-law distribution. A small percentage of your nodes — perhaps 5 to 10 percent — will hold the majority of your connections. This is not a problem to solve. It is a feature of well-organized complex systems.
The reason is mathematical. Preferential attachment produces power-law degree distributions as a natural consequence. Barabasi and Albert proved this: any growing network where new nodes preferentially connect to high-degree nodes will converge on a power-law distribution. Since your knowledge graph grows (you keep adding notes) and new notes preferentially connect to concepts you already think about often (which tend to be the highly connected nodes), your graph will develop hubs whether you intend it or not.
What you can control is how you respond to this structure:
Invest in your hubs. Your top twenty nodes by connection count are the load-bearing infrastructure of your thinking. Review them quarterly. Rewrite them when your understanding deepens. Split them when they grow too broad — a hub that tries to cover too much becomes vague, and vague hubs weaken every connection that passes through them.
Watch for emerging hubs. A node that went from three connections to twelve in the past month is becoming structurally important. It deserves to be well-written before it accumulates more dependents. Upgrade it now, while you still remember the nuance.
Distinguish real hubs from artificial ones. A concept like "feedback loops" is a real hub — it genuinely connects to dozens of domains because the concept itself is cross-domain. A note called "Miscellaneous Ideas" that links to fifty unrelated notes is an artificial hub that adds no structural value. One strengthens your graph. The other obscures it.
Accept the inequality. Most of your notes will have few connections. That is fine. Not every thought is a load-bearing concept. Peripheral notes have value as specific instances, examples, or raw material. But they do not need the same maintenance investment as the nodes that hold everything together.
From structure to strategy
Understanding hub nodes transforms knowledge management from a flat activity ("review all your notes periodically") into a strategic one ("invest disproportionately in the nodes that hold your graph together"). The Pareto principle applies: roughly 20% of your nodes carry 80% of your graph's structural integrity. Maintaining those nodes well is the highest-leverage activity in your knowledge practice.
This lesson connects directly to what comes next. If hub nodes are the most connected nodes in your graph, bridge nodes — the topic of the next lesson — are the most strategically connected. A hub has many connections within and across domains. A bridge node has fewer connections, but the ones it has link otherwise disconnected clusters. Both are high-value. Both deserve your attention. But they serve different structural roles, and understanding the difference is how you move from maintaining a graph to engineering one.