Daniel sent us this one, and I've been turning it over in my head all morning. He's asking: if we had to map the human brain's memory systems onto computer architecture — short-term memory, long-term memory, the way we structure and retrieve memories — what would be the closest technical analogies? Not the pop-science "your brain is a hard drive" stuff, but the real parallels. Where do the metaphors actually hold up, and where do they break?
Oh, this is a beautiful question. And by the way, today's episode is being written by DeepSeek V four Pro, so we'll see if it can keep up with what I'm about to unload.
Bold claim from a donkey who once spent fifteen minutes explaining RAM timings to a pizza delivery guy.
He was building a PC!
He asked if you wanted extra pepperoni, Herman. That's not the same thing.
Look, the point is, this question sits right at the intersection of neuroscience and computer architecture, and most people get it wrong in exactly the same way. They reach for the hard drive analogy and stop there. Long-term memory is the hard drive, short-term memory is RAM, done. But that's not even close to what's actually happening.
Alright, walk me through it. Where does the hard drive analogy fall apart first?
In about six different places simultaneously, which is itself a very brain-like thing to do. Let's start with the most fundamental difference. In a computer, when you store something to a hard drive, you're storing a literal copy. The bits that went in are the bits that come out. The brain doesn't store anything.
Reconstructs from what?
From fragments distributed across the cortex. When you remember your seventh birthday party, there isn't a file called "seventh birthday dot mem" sitting in your hippocampus waiting to be opened. What actually happens is your brain reactivates the pattern of neural firing that occurred during the original experience. It's a replay, not a retrieval.
It's more like... a save state in an emulator? You're not storing the game, you're storing the exact configuration of the system at that moment?
That's actually not a bad starting point, but it's still too clean. The brain's memory is content-addressable, not location-addressable. In a computer, if you want a specific file, you need to know where it lives — the path, the address, the pointer. In the brain, you access memory by providing a partial cue, and the network completes the pattern.
Give me an example of content-addressable versus location-addressable in practice.
Say I ask you to remember what you had for breakfast three Tuesdays ago. You probably can't. But if I say "three Tuesdays ago, you were running late, you grabbed something from that place on the corner, the one with the green awning," suddenly the whole memory snaps into place. The partial cue — green awning, running late — was enough to trigger pattern completion across the network. That's content-addressable memory. You didn't need the memory address "breakfast log entry number four hundred and twelve." You needed a fragment of the content itself.
The brain indexes by content, not by address. That sounds wildly inefficient from a computer science perspective.
It's horrendously inefficient if you're counting operations, but it's incredibly robust. A computer with a corrupted memory address loses the whole file. The brain degrades gracefully — lose a few neurons, the memory gets fuzzier but doesn't disappear. This is actually why the closest computer science analogy for long-term memory isn't a database or a file system. It's a model.
As in, a trained neural network.
When you train a neural network on a dataset, you're not storing the training examples anywhere in the final model. The model's weights encode statistical regularities that allow it to reconstruct or generate outputs similar to what it was trained on. That's what your brain is doing. Your memories aren't stored — they're latent representations that can be regenerated given the right prompt.
Every time I remember something, I'm prompting my own internal model and it's generating a plausible reconstruction.
Yes, and this is where it gets genuinely unsettling. Every time you retrieve a memory, you're not pulling it from a shelf. You're regenerating it, and that regeneration process makes the memory labile — malleable — again. It has to be reconsolidated, essentially re-stored. Which means every act of remembering is also an act of subtle rewriting.
That explains a lot about family arguments. We're all running different fine-tuned models trained on overlapping but distinct datasets.
Each retelling is another fine-tuning step. By the time you've told a story ten times, you're not remembering the original event. You're remembering your last retelling of it.
Alright, so long-term memory as a generative model with latent representations. That's a much better analogy than a hard drive. What about short-term memory? That's the one people always map to RAM.
RAM is actually a reasonable starting point for working memory, but only if you understand what kind of RAM we're talking about. Working memory isn't just a temporary buffer. It's more like a very small register file with active maintenance requirements.
Break that down.
Classic working memory research — Baddeley and Hitch, going back to the seventies — suggests we can hold roughly four to seven chunks of information in working memory at once. That's tiny compared to even the cheapest RAM stick. But more importantly, working memory isn't passive storage. If I give you a phone number and you don't rehearse it, it's gone in about eighteen to thirty seconds. The information decays unless you actively refresh it.
It's more like DRAM that needs constant refreshing?
That's exactly the parallel. DRAM cells leak charge and need to be refreshed every few milliseconds or they lose their state. Working memory representations decay unless they're actively maintained through attentional rehearsal. The mechanism is totally different — neurons doing oscillatory firing patterns versus capacitors leaking electrons — but the functional constraint is strikingly similar.
What's doing the refreshing? Is there a memory controller equivalent?
That's the prefrontal cortex, and this is where the analogy gets richer. The prefrontal cortex doesn't just hold information — it selectively attends to what's relevant and suppresses what isn't. It's the memory controller plus the CPU's scheduling logic rolled into one. When you're trying to remember a phone number while someone's shouting numbers at you, your prefrontal cortex is running interference, actively protecting the representation from degradation.
We've got a two-tier system. Working memory as a small, actively-maintained register file in the prefrontal cortex, and long-term memory as a distributed generative model across the cortex. But there's a piece missing. How does information actually move from one to the other?
This is where the hippocampus comes in, and it's maybe the most interesting architectural parallel in the whole brain. The hippocampus acts as a kind of index server.
An index server. Not the storage itself.
The hippocampus doesn't store your memories. What it stores is a sparse index — essentially a set of pointers to the cortical patterns that were active during an experience. When you have an experience, the sensory details are processed all over your cortex. Visual details in visual areas, sounds in auditory areas, emotional valence in the amygdala. The hippocampus binds all these distributed activations together into a single episodic index.
The hippocampus is maintaining a lookup table that maps episode IDs to distributed cortical patterns.
Yes, and it's not even a permanent lookup table. Over time, as memories are replayed and consolidated — primarily during sleep — the cortical connections strengthen to the point where they can reconstruct the memory without hippocampal involvement. The index becomes redundant, and the hippocampus can let go.
a cache eviction policy. The hippocampus caches recent memories, then gradually offloads them to long-term cortical storage once they're sufficiently consolidated.
You're laughing, but that's exactly what's happening. Researchers call it systems consolidation. The hippocampus is critical for recent episodic memories but becomes less necessary over time. Patients with hippocampal damage can't form new memories but can often recall old ones just fine. The old memories have been fully offloaded to the cortex.
If I'm mapping this to a modern system architecture, I'm seeing: working memory as L one cache — tiny, fast, actively maintained. The hippocampus as an indexing layer with a temporal decay function — it prioritizes recent, unconsolidated memories. And the cortex as a massive distributed storage substrate that stores latent representations, not literal data.
The consolidation process during sleep is essentially an offline batch re-indexing job. The brain replays experiences at roughly twenty times speed, with the hippocampus repeatedly reactivating cortical patterns until the cortex can self-activate them.
Twenty times speed. So my brain is running a nightly ETL pipeline on my daily experiences.
With selective retention heuristics. Not everything gets consolidated. The brain prioritizes emotionally salient experiences and information that connects to existing knowledge. It's doing deduplication and relevance scoring.
This is starting to sound less like a storage system and more like a retrieval-augmented generation pipeline. Which, given what Daniel works on, is probably why he asked this question.
The RAG parallel is actually really tight. In retrieval-augmented generation, you've got a language model that doesn't store all its knowledge in its weights. When it needs to answer a question, it queries an external knowledge base, retrieves relevant documents, and uses those to ground its generation. The brain does something remarkably similar.
Walk me through the brain's RAG pipeline.
When you're asked a question — say, "what did you have for dinner last night" — your prefrontal cortex formulates a query. That query activates partial patterns in your cortex, which serve as retrieval cues. The hippocampus, acting as the retrieval engine, uses those cues to reactivate the full episodic pattern from the relevant cortical areas. That reactivated pattern is then held in working memory, where your prefrontal cortex can use it to generate a response.
The query formulation, retrieval, and generation are all happening in the same substrate, just in different regions. There's no separate vector database sitting off to the side.
That depends on how you think about it. The cortex is both the vector database and the generation model. The representations that store memories are the same representations that process incoming sensory information and generate motor outputs. There's no clean separation between storage and compute.
That's a von Neumann architecture violation right there. In a standard computer, memory and processing are physically separate. The brain is doing compute-in-memory by default.
Every neuron is simultaneously a storage unit and a processing unit. The strength of a synapse — its weight — stores information, but that same synapse is also performing computation every time it fires. It's closer to a neuromorphic chip architecture than anything in a standard server rack.
Alright, let's get concrete for a minute. If I'm a software engineer and I want to build something that mimics this architecture, what am I actually building? Give me the system diagram.
You're building a system with three main components. First, a working memory module — very small capacity, maybe four to seven item slots, with active maintenance through recurrent connections. This is basically a recurrent neural network with a very limited hidden state that decays unless refreshed.
Four to seven slots. That's absurdly small.
It is, and there's a fascinating reason for it. If working memory were much larger, the combinatorial explosion of possible associations would make retrieval noisy and error-prone. The capacity limitation is actually a feature, not a bug. It forces the brain to be selective about what it holds onto.
Alright, component one is the working memory buffer. What's component two?
Component two is the hippocampal index. This is a fast-learning system that can form new associations in a single exposure — what computer scientists would call one-shot learning. The hippocampus uses a specialized circuit architecture called the trisynaptic pathway that allows it to rapidly encode novel patterns without overwriting previous ones.
That's the catastrophic forgetting problem that plagues standard neural networks.
Standard neural networks tend to overwrite old knowledge when learning new information. The brain solves this through complementary learning systems — the hippocampus learns fast and acts as a temporary buffer, while the cortex learns slowly and integrates new information gradually through interleaved replay.
The hippocampus is essentially doing online learning with a high learning rate, and the cortex is doing offline fine-tuning with a much lower learning rate.
That's remarkably accurate. The hippocampus has a learning rate that's orders of magnitude faster than the cortex. It can form a new association in seconds, but it's also more prone to interference. The cortex learns slowly but builds robust, generalizable representations.
Component three is the cortical storage system — a massive, distributed, content-addressable memory implemented as a hierarchical generative model. Each cortical area learns to predict the activity patterns in the areas below it, building increasingly abstract representations as you move up the hierarchy.
So visual cortex has low-level edge detectors, then shape detectors, then object representations, then conceptual representations that link objects to semantic knowledge.
Each level can be queried bidirectionally. You can ask "what does a cat look like" and the system generates a visual representation from the top down. Or you can see a cat and the system activates the concept from the bottom up. It's a bidirectional generative model.
This is starting to sound like a very specific architecture that's been getting attention lately. The hierarchical bidirectional stuff, the content-addressable memory, the complementary learning systems...
You're thinking of the thousand brains theory from Jeff Hawkins and Numenta, aren't you?
I was, but I didn't want to steal your thunder.
Hawkins' framework maps onto this really well. His core argument is that the neocortex is organized into roughly one hundred fifty thousand cortical columns, each of which is running essentially the same algorithm. Every column builds a model of the world from its specific sensory perspective, and they vote — literally vote — on what the overall percept is.
It's a distributed consensus protocol. Each cortical column is a node proposing a hypothesis about what it's sensing, and the winning hypothesis emerges from collective voting.
The voting is continuous and real-time, happening in milliseconds. It's like a blockchain consensus mechanism but actually fast and useful.
I'm going to pretend you didn't just say blockchain in a neuroscience discussion.
The point stands. Every cortical column maintains its own model, its own set of learned representations, and they coordinate through lateral connections. When you recognize an object, you're not running a single classifier. You're running a distributed election across thousands of mini-models, and the coherence of their predictions is what gives you confidence in the percept.
That explains some interesting failure modes. Optical illusions would be cases where the columns reach a split decision or the wrong consensus.
And the tip-of-the-tongue phenomenon — where you know you know a word but can't retrieve it — maps onto a partial retrieval where enough columns are activated to give you a sense of familiarity but not enough to complete the pattern.
The quorum isn't reached.
The quorum isn't reached. You've got the semantic features activated — you know it's a word for that thing, you know it starts with a certain letter — but the full phonological representation can't be reconstructed.
Let's talk about the retrieval process itself, because this is where I think the computer analogies get most interesting. In a database, retrieval is deterministic. You query, you get the exact record or you don't. In the brain, retrieval is probabilistic and reconstructive. What's the closest technical analog for that?
The closest analog is probably a variational autoencoder — a VAE. In a VAE, you don't store exact representations. You store probability distributions over latent variables. When you want to reconstruct something, you sample from that distribution, which means every reconstruction is slightly different.
My memory of my wedding is a probability distribution, and every time I recall it, I'm sampling a slightly different version.
The distribution itself shifts over time. The mean might drift, the variance might increase. Details that were once crisp become fuzzy not because they're "deleted" but because the probability distribution becomes less peaked.
That's a much more elegant explanation for memory degradation than "the file got corrupted." The distribution is flattening.
It also explains why memories can be simultaneously unreliable and stable. The gist of a memory — the central tendency of the distribution — can remain stable for decades. But the specific details — the tail of the distribution — become increasingly unreliable.
There's something else that's been nagging at me. In a computer, when you store a piece of data, you have a specific write operation. You receive the data, you write it to storage, done. But the brain doesn't seem to have a clean separation between encoding and retrieval. They're intertwined.
That's one of the most important differences, and it's rarely discussed. In the brain, the same circuits that process information are the ones that store it. When you perceive something, the pattern of neural activity IS the encoding. There's no separate "save" operation. The act of experiencing is the act of encoding.
Which means attention is the brain's write-enable signal. If you're not paying attention, the pattern never forms strongly enough to be consolidated.
This is why multitasking destroys memory formation. If your prefrontal cortex is dividing attention across multiple streams, none of them get the full encoding signal. The patterns are weak, fragmented, and unlikely to survive consolidation.
The brain's write policy is essentially "write-through with attention-gated selectivity." Data passes through working memory, and only the attended, emotionally-tagged, sleep-consolidated subset makes it to long-term storage.
With an important caveat. Even the stuff that doesn't make it to explicit, episodic memory still leaves traces. Procedural memory, priming effects, implicit learning — these all happen through different mechanisms that don't require conscious attention.
The distinction between declarative and procedural memory. "Knowing that" versus "knowing how." Are those different storage systems entirely?
Different neural substrates, yes. Declarative memory — facts, events, explicit knowledge — relies heavily on the hippocampus and medial temporal lobe structures we've been discussing. Procedural memory — skills, habits, conditioned responses — relies more on the basal ganglia and cerebellum.
If declarative memory is a generative model with a hippocampal index, what's procedural memory?
Procedural memory is closer to a finely-tuned control policy. It's not about reconstructing past experiences. It's about optimizing future actions. When you learn to ride a bike, you're not storing a memory of riding a bike. You're tuning a feedback control system that maps sensory inputs to motor outputs.
That sounds like reinforcement learning. The basal ganglia are running something like a policy gradient algorithm.
The dopamine system in the basal ganglia implements something remarkably similar to temporal difference learning — the same algorithm at the heart of modern reinforcement learning. When something turns out better than expected, dopamine neurons fire more. When it's worse than expected, they fire less. This prediction error signal drives learning.
We've got supervised learning for episodic memory consolidation, reinforcement learning for skill acquisition, and unsupervised learning for building cortical representations. The brain is running multiple learning algorithms in parallel on different substrates.
Your declarative knowledge about how to swing a golf club can influence your procedural learning of the swing. Your procedural habits can bias what you pay attention to and therefore what gets encoded into episodic memory. The systems aren't isolated.
Let me try to synthesize this into something coherent. If I were to describe the brain's memory architecture to a senior engineer who's never studied neuroscience, I'd say something like this.
Go for it. I'll critique.
The brain implements a multi-tier memory architecture with no separation between storage and compute. Working memory is a small register file of four to seven slots, actively maintained through recurrent dynamics in the prefrontal cortex — analogous to DRAM with an attentional refresh cycle. Long-term declarative memory is a distributed, content-addressable storage system where memories are stored as latent representations — probability distributions over neural activation patterns — rather than literal records. A specialized structure called the hippocampus serves as a fast-learning index server, binding distributed cortical patterns into coherent episodic indices. Over time, through sleep-based replay and consolidation, these indices become redundant as cortical connections strengthen, effectively implementing a cache eviction policy where the hippocampus offloads consolidated memories to the cortex.
Procedural memory operates on a separate substrate using reinforcement learning principles, with the basal ganglia implementing something akin to temporal difference learning for skill acquisition. The entire system is bidirectional and hierarchical, with each cortical level generating predictions about the levels below it. Retrieval is probabilistic and reconstructive — more like sampling from a variational autoencoder than querying a database. And the whole thing runs on roughly twenty watts.
That's the number that should humble every computer architect. The human brain does all of this — perception, memory, reasoning, motor control, language — on about twenty watts of power. A single high-end GPU draws ten to twenty times that just to run a language model.
Though to be fair, the brain had a few hundred million years of evolutionary hyperparameter tuning.
It's running on wetware that operates at millisecond timescales, not nanosecond transistor switching. The fact that it's competitive at all is remarkable. The fact that it still outperforms silicon on most general intelligence metrics is staggering.
There's one more analogy I want to explore, and it connects back to something Daniel works with regularly. The retrieval process in the brain has a strong similarity to what happens in modern RAG systems, but with a twist. In RAG, you've got a clear separation between the retriever and the generator. The retriever fetches documents, the generator uses them. In the brain, those aren't separate modules.
Right, and this is where the brain's architecture is arguably more elegant. The same cortical representations that serve as the "documents" in the knowledge base are also the representations the "generator" uses to produce output. When you retrieve a memory, you're not fetching a document and feeding it to a separate language model. The act of retrieval IS the act of reconstruction. The memory and the process that reads the memory are the same thing.
That eliminates the context window problem entirely. In an LLM with RAG, you're limited by how many retrieved documents you can fit in the context window. The brain doesn't have that bottleneck because retrieval and generation happen in the same representational space.
It has a different bottleneck. The four-to-seven chunk working memory limit is the brain's context window. It's tiny by LLM standards — a few tokens effectively, compared to hundreds of thousands for modern models. But the brain compensates through hierarchical chunking.
Explain chunking in this context.
A chunk is a unit of information that's been bound together through learning so it occupies only one slot in working memory. The classic example is chess masters. Show a chess master a board position for five seconds, and they can reconstruct it almost perfectly. But only if the position is from a real game. If the pieces are arranged randomly, their performance drops to novice level.
Because the real game positions form meaningful chunks. A particular pawn structure, a common opening configuration — those get encoded as single units.
The chess master doesn't see twenty-five individual pieces. They see four or five chunks, each representing a meaningful configuration. That fits in working memory. The random board can't be chunked, so it exceeds capacity.
Chunking is essentially learned compression. You're building higher-level tokens that represent common patterns.
It's hierarchical all the way up. Letters become words, words become phrases, phrases become ideas. A skilled writer isn't thinking about individual letters when composing a sentence. They're manipulating idea-level chunks, and the lower levels unfold automatically.
This connects to something I've been thinking about with language models. The context window in an LLM is flat — every token costs the same in terms of attention computation. But the brain's working memory is deeply hierarchical. You pay attention to the highest level of abstraction that's relevant, and the details fill in through learned associations.
There's active research on hierarchical attention mechanisms that try to mimic this. Instead of attending to every token equally, you learn to attend to chunks at multiple levels of abstraction. It's more efficient and arguably more human-like.
Alright, let's talk about where the analogies actually break. You've been generous to the computer science parallels so far. Where does the brain do things that have no good computational analog?
The biggest one, and I don't think this gets enough attention, is the role of emotion in memory encoding and retrieval. In a computer, all bits are equal. A wedding photo and a picture of your grocery list get the same storage treatment. In the brain, emotional salience acts as a gain control on encoding strength.
The amygdala is modulating the hippocampal consolidation process.
Norepinephrine and cortisol, released during emotionally arousing events, directly enhance the molecular processes that strengthen synaptic connections. A traumatic event gets burned in. A boring Tuesday doesn't. There's no computational reason to treat those differently in a pure information storage system, but the brain does it automatically.
That's a feature, not a bug, from a survival perspective. Remembering where the predator lives is more important than remembering what shade of brown the dirt was.
It creates systematic biases in memory that are hard to model computationally. Flashbulb memories — where you vividly remember where you were during a major event — feel incredibly detailed and accurate, but research shows they're just as prone to distortion as ordinary memories. They just come with higher subjective confidence.
The emotional tagging doesn't improve accuracy. It improves persistence and subjective vividness.
Which are different things. And that distinction gets lost in most computer analogies.
Another thing that seems hard to map: the brain does massive amounts of offline processing during sleep that we barely understand. Memory consolidation, certainly, but also something that looks like generalization and abstraction.
The replay during sleep isn't just verbatim replay. The hippocampus replays experiences in compressed form, but it also interleaves fragments from different experiences, and it sometimes generates novel combinations. This is thought to underlie insight and creative problem-solving.
The nightly batch job isn't just backup. It's doing data augmentation.
Potentially counterfactual reasoning. By recombining elements from different experiences, the brain can explore "what if" scenarios without the cost of actually living through them. That's a capability that current AI systems don't really have in the same integrated way.
I want to push on one more thing. We've been talking about memory as if it's a single, coherent system. But the brain seems to have multiple memory systems that sometimes conflict. How do you model that?
You're talking about cases like the Muller-Lyer illusion, where you consciously know the lines are the same length, but you can't stop perceiving them as different. That's a conflict between your perceptual system's implicit memory — which has learned that certain visual configurations indicate depth and therefore different sizes — and your explicit knowledge that they're identical.
You've got two memory systems producing contradictory outputs, and the explicit system can't override the implicit one.
This is totally alien to computer architecture. In a computer, if two processes produce conflicting results, one of them is buggy and you fix it. In the brain, conflicting outputs from different systems is the normal operating state. Your prefrontal cortex is constantly adjudicating between competing impulses, perceptions, and memory traces.
It's more like a parliament than a pipeline.
A very noisy parliament with no clear majority party and constant filibustering from the amygdala.
That explains approximately all of human history.
The point is, the brain isn't a unified architecture designed by a single engineer with a coherent plan. It's a stack of systems that evolved at different times for different purposes, jury-rigged together and constantly negotiating with each other. Any computer analogy that makes it sound clean and elegant is missing the fundamental messiness.
If you were advising someone building AI systems inspired by the brain, what would you tell them to steal and what would you tell them to ignore?
Steal the complementary learning systems idea — fast online learning paired with slow offline consolidation. That's useful and under-explored in production systems. Steal the content-addressable, distributed representation approach. Steal the hierarchical predictive processing framework — the idea that every level of the system is trying to predict the level below it, and what gets passed up is only the prediction error.
Ignore the messiness. Ignore the emotional modulation unless you're building something that needs affective computing. And definitely ignore the power consumption constraints — twenty watts is aspirational, not practical, for silicon.
One last question. Given everything we've discussed, what's the single most misleading thing people say about brain memory?
That memory is like a video recording. It's not. It's not even close. Every time you remember something, you're reconstructing it from fragments, filling in gaps with plausible inferences, and subtly altering it in the process. Your memories are more like stories you tell yourself — ones that get revised with each retelling.
Yet they feel so real.
That's the most impressive feature of the whole system. The generative model is so good at reconstructing coherent narratives that you don't notice the reconstruction process. It feels like playback. It's not.
I think we've earned our fun fact.
Now: Hilbert's daily fun fact.
Hilbert: In the nineteen twenties, researchers estimated that a single ant colony on Kiritimati Island in Kiribati produced enough trail pheromones in one year to mark a continuous path stretching from the island to Fiji — a distance later measured at roughly two thousand kilometers — with chemical concentrations so precise that individual ants could discriminate between their colony's trail and a neighbor's at concentrations as low as one part per trillion.
...right.
Ants can navigate by chemical concentration gradients at the parts-per-trillion level, and I still can't find my keys.
To be fair, your keys aren't emitting pheromones.
They might be. I haven't checked.
So where does all of this leave us? The brain as a multi-tier, content-addressable, distributed generative model with a hippocampal index, a tiny working memory buffer, and a twenty-watt power budget. It's simultaneously the most elegant and the most jury-rigged information processing system we know of.
The computer analogies are useful but limited. The brain isn't a computer. But thinking about it in computational terms reveals just how many clever architectural tricks evolution stumbled onto — complementary learning rates, hierarchical chunking, probabilistic reconstruction, compute-in-memory. These aren't just biological curiosities. They're design principles that are increasingly showing up in the AI systems Daniel and his colleagues are building.
Which is probably why he asked the question in the first place. Thanks to our producer Hilbert Flumingtop, and thanks to Daniel for another great prompt.
This has been My Weird Prompts. You can find every episode at myweirdprompts dot com. If you enjoyed this, leave us a review wherever you listen — it helps.
Until next time.
I'm going to go see if my keys emit pheromones.
