Some mornings I wake up having composed something extraordinary. I mean it. In the dream there was a full piece, arranged, moving, better than anything I have written with my hands, and for a few seconds after waking it is still playing. Then it does what dream music does. It thins, it slips, and by the time I have found something to hum into, there is a vague contour left and a very precise feeling of having been robbed. Not by anyone. By the distance between my ears and my hands.
The daytime version is quieter but just as maddening. As a musician (an amateur one, several instruments, none of them well), I hear ideas in my head all day, and they fail in two ways. Either they are too fleeting to grab, gone before I have found a keyboard. Or they are too complex. The moment I focus on one voice to write it down, the others collapse behind it, like turning to look at one bird and losing the flock.
The worst part is not forgetting. It is remembering that it was good.
So here is the device, the fifth climb up this site’s science fiction ladder (after the storm, the gun, the collar and the air), and the third about music, counting the volume knob. It is a recorder for the inner ear. Something that captures the music you imagine, as you imagine it, whole, without passing through your mouth or your fingers. Not a dream recorder. Dreams are another essay, with a thousand years of literature attached. Just this one narrow, ancient, ridiculous need: to hear a thing in your head and keep it.
Three Centuries of the Same Morning
Ancient is the right word, because the problem has a founding document. In 1713, the violinist Giuseppe Tartini dreamed that the devil sat at the foot of his bed and played him a sonata of unbearable beauty. He woke, grabbed his violin, and wrote down what he could. The result is the “Devil’s Trill,” one of the most demanding pieces in the repertoire, and Tartini said for the rest of his life that it was infinitely inferior to what he had heard. He had my morning, in 1713, and he had it with better material.
He has company. Paul McCartney woke up with the melody of “Yesterday” complete in his head, went to the piano to trap it, and then spent weeks asking everyone he knew whether it already belonged to someone else, because nothing that arrives that finished feels like yours. And the most instructive case is Keith Richards, in 1965, who caught the riff of “Satisfaction” half-asleep on a bedside cassette recorder. In the morning the tape held about two minutes of the riff, followed by forty minutes of snoring. The bedside recorder is the ancestor of the device I want. It worked exactly once in the history of music, and it is still the best tool we have.
The losses have their monument too. Coleridge composed “Kubla Khan” in a dream, woke, began writing it down, and was interrupted by a visitor from Porlock. When he sat back down the rest was gone, and the poem has been published ever since as a fragment. Every musician I know has a person from Porlock. Mine is usually an alarm clock.
The Bottleneck Is the Mouth
Why does humming fail? Because the inner ear does not hold music as a sequence. It holds it as a shape. The chords, the bass moving under them, the drums, and the line on top are all present at once, the way you hold a face. But the only exits available are serial. A mouth can carry one line at a time. A hand on a keyboard can carry a few, slowly. And serializing the shape destroys it. By the time you have hummed the melody, the harmony that made it worth humming has evaporated, and you are left with a tune that sounds, correctly, like a thinner version of a thing you can no longer compare it to.
Today’s tools inherit the bottleneck. Voice-to-MIDI and hum-to-notation apps exist and work well, one monophonic line at a time, which is exactly the constraint the problem is made of. A phone’s voice memo is the Richards method with better tape. None of them touch the actual fault, which is that a rich musical thought cannot survive being pushed through a hole the width of one note.
This is the deep reason the device deserves to exist rather than a better dictaphone. Reading the inner ear directly would read it in parallel, the whole flock rather than one bird at a time, and parallel is the only way the shape survives the trip.
What the Science Just Did
What I did not expect when I started checking was that the first rung of this ladder had been climbed weeks before I wrote this. A study published in 2026 in eNeuro decoded imagined melodies of familiar children’s songs from ECoG recordings in ten epilepsy patients, by representing melody as relative pitch (do, re, mi rather than absolute frequencies), and the sequence of decoded notes reconstructed the contour the patients were imagining. Single-note accuracy was modest but reliably above chance. Imagined music, read from a brain, as a contour on a screen.
The non-invasive side has been creeping up too. When you imagine music, your brain produces a response that follows the amplitude envelope of the imagined sound, and researchers have decoded that envelope from ordinary scalp EEG well enough to recognize which song a person is imagining. Another team asked trained musicians to imagine four Bach melodies and decoded them, bar by bar, from EEG alone. And on the perception side, a 2023 study reconstructed a recognizable Pink Floyd song straight from patients’ cortical recordings, while others have rebuilt heard music from brain scans by pairing them with generative audio models.
Now read the small words in those sentences: familiar, bar by bar, contour, recognize. Every one of these systems works on music that already exists. The patients imagine the continuation of a children’s song they know. The musicians imagine Bach they have learned. The decoders have a template to match against. A melody that arrived in your sleep last night has no template anywhere in the world. That is the distance left to climb.
So the device is a stack of four leaps, each one real on its own, fictional only together. The first goes from known to novel, decoding music nobody has heard yet, with no reference recording to lean on. The second goes from one voice to many. Today’s proofs of concept follow a single melodic line, but my dreams arrive arranged. The third goes from electrodes on the cortex to something you can wear. The good results come from surgery patients, scalp EEG is coarse, and a brain scanner is not a bedside object. The last goes from waking to the edge of sleep, where the best material lives and the signal is worst.
I want to be exact about which rung is science fiction, because the answer is unusually flattering to the dream. None of them is, individually. Novel-content decoding is where imagined-speech research is already heading. Polyphony is a bandwidth problem, not a physics problem. Wearable neural sensing is an entire industry’s roadmap. The fiction is only the stack, worn on one head, at 6am. Most real devices were fiction in the same way, right up until they were not.
The Edge, Not the Dream
Where would you wear it? Not, I think, all night. Dream recording is someone else’s ambition, and a heavier one. What matters for music is the edge: the hypnagogic minutes falling asleep, when ideas arrive unbidden, and the hypnopompic minutes waking up, when the dream piece is still faintly playing and already leaving. Two anchors make that edge a real place rather than a poetic one. In 2021, a team at the Paris Brain Institute recreated Thomas Edison’s famous nap trick under laboratory conditions and showed, in Science Advances, that the very first stage of sleep is a measurable hotspot for creative insight. And the MIT Media Lab built Dormio, a glove that detects the moment you drift off and gently prompts you, capturing the thoughts of the edge before they dissolve.
My device borrows the glove and gives it an ear. It knows when you are waking before you do, and it opens the capture in that window, before the eyes, before the alarm, before the person from Porlock. Nothing to press, nothing to find, no forty minutes of snoring. The Richards method with a nervous system attached.
Was It Ever Notes?
And now the question I cannot answer, which is why this is an essay and not a product. The most impressive decoders today do not read music out of a brain so much as generate plausible music conditioned on it: a model’s best guess at what would produce those signals. Point one at a dream and it will hand you something confident, competent, and possibly not yours. Would you know? You would have nothing to compare it to except a feeling.
That raises the harder possibility. Maybe the dream piece was never notes. The sleeping brain manufactures the sensation of significance on demand. Everyone has woken from a dream certain they had understood something enormous, holding nothing. Perhaps the phenomenal symphony is, at least sometimes, the feeling of music rather than music, and Tartini’s lifelong disappointment was not a failure of memory but a rare honest glimpse of the gap. The device I want could confirm my nocturnal genius. It could also, on its very first morning, break the spell for good.
I would build it anyway. Here is one fragment you could build this afternoon, in the spirit of this site. A sleep-tracking wearable already knows when you wake, and a phone can open a one-tap recorder in that second, then walk you through the flock one bird at a time (“melody first; now the bass; now the drums”) before the shape is gone. It is the cassette recorder with a nervous system, and it would rescue more than you think. If you build it, tell me, and I will send you my first recording. It will probably be forty minutes of snoring.
And if Tutti ever gets its dark stage, this is the music that fills it: the phrase caught at 6am, shaped by the same hands that failed to hum it.
So here is what I keep turning over. If you could finally hear, in daylight, the music your sleeping mind is so proud of, would you want to know how good it really was?
