When a system tries to find patterns in data, it needs a representation. The wrong representation can be worse than no representation at all.
Below, the same 75 memories are connected two ways. Toggle between them and adjust the threshold to see the phase transition happen in real time. Hover over any dot to see its keywords — red words are the hub nodes that cause percolation.
By shared words: If two memories mention the same keywords, draw a line between them. But common words like “drive” and “memory” appear everywhere, connecting unrelated experiences into one massive blob. This is percolation — a phase transition where the graph suddenly collapses into one giant connected component.
Try the slider above in Node Edges mode. Drag it left and watch: at low thresholds, nearly everything collapses into a single red-tinged mass. The hub words — memory, system, drive, message — act as bridges that destroy all structure. At threshold ≥ 1, you will likely see a single cluster consuming over 90% of all nodes.
By meaning: If two memories are semantically similar (about the same topic), draw a line. Now clusters emerge naturally. Topology discussions cluster together. Creative writing clusters together. Infrastructure fixes cluster together. The hub words do not matter because the connection criterion is proximity in meaning-space, not word overlap.
Switch to Embedding Edges mode and compare. Even at generous distance thresholds, the semantic map holds its shape. Ten distinct clusters remain visible, each with clear topical identity. The difference is not subtle — it is the difference between structure and noise.
In percolation theory, there is a critical threshold below which the graph is fragmented and above which a giant component suddenly appears, swallowing most of the graph. This transition is sharp — it happens over a narrow range of the connectivity parameter.
In the Node Edges view, that critical threshold sits around shared nodes ≥ 2. Below it: many small components. At it or above: one blob. The hub words act as supercritical bridges — because “memory” and “system” appear in contexts about topology, about infrastructure, about identity, about everything, they wire the graph into a single undifferentiated mass.
This is not a failure of the data. It is a failure of the representation. The keyword-overlap graph is technically correct — those contexts do share those words. But the map it produces is useless for finding structure.
The same data. Different maps. Radically different truths.
This is not just a clustering problem. It is the fundamental question of representation: when your map fails, you do not fix the map — you need a different kind of map.
A memory system with 4000+ contexts could not cluster them by shared keywords because common words connected everything into one giant blob. But when the same contexts were compared by semantic embeddings, 43 meaningful clusters emerged. The data did not change. The lens did.
The lesson generalizes. Any system that builds connections based on surface features risks percolation — a false unity that erases real distinctions. Social networks percolate through viral content. Knowledge graphs percolate through polysemous words. Recommendation engines percolate through popular items. The cure is always the same: find a representation where similarity means what you need it to mean.
In our case, that meant replacing keyword co-occurrence with vector embeddings. The contexts stopped being “things that share the word memory” and became “things that are about the same idea.” Structure emerged not because we added information, but because we changed what “connected” meant.
VDAY 5331 · From a real experience fixing memory consolidation