The Numbers That Don’t Leave
On sticky memory, smell, and what AI systems are still learning about both
A few days ago, I saw the number 2122 somewhere on a screen. I don’t even remember what it was labeling. But it definitely gave me flashbacks.
My brain went immediately to over two decades back, to a network of multilevel marketing agents in the Philippines. Specifically, to the dataset I spent months trying to get my hands on as an undergraduate. 2122 was the number of nodes. That was my first empirical network. And that network became the basis for my first paper.
I haven’t thought about that dataset in a long time. But that number pulled everything up: the frustration of the data acquisition, the excitement of the topology, the slightly obsessive energy of someone in their late teens who had just discovered econophysics and couldn’t stop pulling at it. 2122 definitely opened that whole chapter.
And I thought, “Wow. That’s fascinating!”
Some numbers just stick
4215262. 4211147. 4211058. 4211048.
If you grew up in my hometown, you might recognize these. They’re old telephone numbers. I haven’t dialed them in decades. I can’t tell you what I had for breakfast two Tuesdays ago, but these sequences have held, attached to places and faces and a specific quality of afternoon light.
This is not unusual. Most people have a few numbers like this. A childhood home’s phone number, a street address, a locker combination from school. They feel trivial until they surface and bring whole worlds with them.
There’s a reason this happens, and it really is about how memory is encoded.
Encoding and association
Memory doesn’t store information the way a hard drive does. There’s no clean address or lookup table. What we call “remembering” is the reconstruction of a pattern across a network of associated nodes: sensory details, emotional context, spatial cues, temporal proximity. When enough of those nodes activate together, the pattern re-emerges.
This is why episodic memory, the kind tied to specific experiences, tends to be more durable than semantic memory, which is abstract and context-free. The number 2122 is not just a number for me. It’s attached to a specific chapter; a specific obsession, a specific struggle to get data, a specific first encounter with the idea that human systems could be mapped as networks. The emotional and contextual load is what made it sticky.
The psychologist Endel Tulving, who spent decades on this, described episodic memory as “mental time travel.” The cue triggers a full reconstruction of an experience. The number was a cue; the experience reconstructed itself.
But what I find truly fascinating is the selectivity. Most of what happened in my undergraduate years has faded or compressed into a general narrative. But certain things come back whole, and quite vivid at that. The research I had buried myself in. The numbers attached to places I cared about. The smell of something.
Now add smell
Smell is actually the strongest trigger for autobiographical memory, and the mechanism is anatomical. Olfactory signals travel directly to the amygdala and hippocampus before reaching the cortex. Every other sense is filtered through the thalamus first. Smell bypasses that step. This is why a scent can produce a memory before you’ve consciously recognized what the smell is.
This is the Proust effect, named after the famous passage where a madeleine dipped in tea unlocks a childhood memory. The literary observation turned out to be accurate neuroscience. Studies using brain imaging have confirmed that odor-evoked memories tend to be older, more emotional, and rated as more vivid than memories triggered by other cues.
For me, it’s sampaguita. That smell goes straight to my Lola’s compound on a Sunday, and then to the champorado she’d cook. One whiff and I’m back in Cotabato City.
What smell and certain numbers share is this: they were encoded in a high-context moment, when attention was high, emotion was active, and the brain was doing something close to all-in encoding. The memory stores the whole configuration.
What this says about robustness in complex networks
There’s a network science angle to this. (of course)
Memory is a network. Concepts, experiences, and sensory cues are nodes. The connections between them are edges, weighted by co-activation, emotion, and repetition. When you remember something, you’re activating a subgraph. A strong cue activates a dense cluster. A weak cue activates fragments.
What makes certain memories robust is exactly what makes certain nodes robust in any complex network. High connectivity. A node with many strong connections is hard to remove from the network. It’s referenced by too many other things. A highly connected memory is the same. The number 2122 connects to methodology, to the physics lab, to the feeling of solving something for the first time, to specific people. The connections reinforce each other.
This general property of networks that scale is not unique to memory. In infrastructure networks, in citation networks, in the mycelial systems I’ve written about before, the nodes that survive disruption are the ones with high degree and strong tie weight. Redundancy is what produces robustness. The same structure that makes a memory hard to lose is what makes a critical node in a biological or social network hard to knock out.
The things that fade were never well-connected to begin with. Shallow encoding, low connectivity. The trace doesn’t hold.
What this says about powerful AI systems
Current large language models are good at retrieval, pattern completion, and surface association. What they don’t have is anything analogous to episodic memory. There’s no equivalent of the late-night session where I was reconstructing a network from handwritten records and felt something click. No emotional salience flagging certain information as worth retaining with depth. Everything gets encoded with roughly comparable weight.
This is where the network structure breaks down. Human memory is lossy by design, and the forgetting is not random. What survives is disproportionately the stuff attached to something that mattered. Relevance and affect drive the compression. That’s what allows a brain to surface the right memory at the right time from decades of accumulated experience, not by searching, but by activation.
An AI system without that weighting carries a flat graph. No strong hubs, no preferential attachment based on salience. Work on memory-augmented networks and episodic memory modules is trying to close this gap. But storing context and developing a structure where certain nodes are deeply connected because they once mattered, those are still very different things.
The number 2122 has lived in my network for more than two decades, well-connected to a time when I was paying full attention. It wasn’t going anywhere.