VISIONSeed
Can auditory cues recreate
or retrieve human states?
- Domain
- Memory and interaction design. Songs and photographs carry whole periods of a person's life, and this project asks whether an interface can return them on demand.
- Role
- Solo, April to September 2026. I designed and ran the study, thirty encounters with 5 people including myself, then designed and built ARC and Seed.
- Question
- Can a song or a photograph bring back the state a person was once in, beyond the memory of it?
- Outcome
- Music brought a period's feeling back more strongly than photographs, 4.53 against 3.80 of 6, while photographs more often returned a scene. That became 4 design rules and ARC, an interface that asks before it follows.
VISIONSeed takes the remembrance, nostalgia and memory a person keeps in a song or a photograph and turns it into an interface they can go back to on demand. The longer ambition is to recreate one such state with sound alone.
Music has a fairly obvious relationship with memory. Play something often enough around a person, a place or a period of your life, and there is a reasonable chance some part of that context stays attached to it. A lover shows you a song, another plays incessantly through university, another happens to occupy the few months where everything goes wrong. Years later the song plays again and something comes back. What exactly comes back is where this starts becoming difficult, since it could be a scene, an atmosphere, a feeling or some sense of the person you were then, and we tend to call all of it “this takes me back”.
VISIONSeed asks whether a sound can bring back the state a person was once in, beyond the memory of it. Early versions pushed this toward restoring a former state outright, with fMRI, EEG and physiological measures considered as evidence. The historical state had never been recorded, so none of them could confirm a return. The project kept the question and changed the instruments: a study of what personal songs and photographs bring back; ARC, an app that follows what a person reports; and Seed, which records a period while it is still being lived.
The archive
Next is my listening archive, built in 3D. Scrolling moves you through it.
Read it as text instead
Forty-one months, heard back
This landscape is built from my own listening history: 51,344 recorded plays between April 2023 and August 2026. Each slab is a month, raised to the number of times I pressed play. As data it is only a count. Walking along it, I can still name the months that changed me. In a pilot trial on myself, one song from August 2023 brought the period back within about three seconds, and with it a version of me I could only describe as “a scared boy”.
- Forty-one months, heard back
This landscape is built from my own listening history: 51,344 recorded plays between April 2023 and August 2026. Each slab is a month, raised to the number of times I pressed play. As data it is only a count. Walking along it, I can still name the months that changed me. In a pilot trial on myself, one song from August 2023 brought the period back within about three seconds, and with it a version of me I could only describe as “a scared boy”.
- How a song holds a year
Memory is retrieved by cues that overlap with the way an experience was stored, a principle Endel Tulving called encoding specificity. A song heard constantly within one season, and rarely outside it, becomes an unusually precise cue for that season. In my own archive, some tracks place more than three-quarters of their plays inside a single period. Played again, such a song reinstates part of the context it was learned in.
Ms(t) = sim(Xt, Rs) − maxj≠s sim(Xt, Rj)
A self is a regionIf a former self can return, what exactly returns? A state is never one fixed pattern. It is a region that brain, body and behaviour move within, and neighbouring regions overlap: longing and loneliness can feel alike and still be different places. So I defined return as a margin: how similar the present moment is to the state I am aiming for, minus its similarity to the closest rival. A song only counts if it moves that margin toward the right past.
Xt is the pattern measured after the cue; Rs is a reference recorded for the target state; Rj are the rival states. Becoming sadder, calmer or more nostalgic does not raise the margin unless the pattern moves toward the right state and away from its neighbours.
- The self nobody recorded
That margin needs a reference recorded during the original period, and nobody measures their life while living it. I designed the experiment that could settle the question anyway, then accepted what the design exposed: the past selves I most wanted to return to were never recorded, so their return can never be verified. What can be studied honestly is what people report when a cue reaches them. That became the experiment.
Thirty encounters
Five people, ten periods, thirty seconds each
Four people and I each chose two periods of our lives, which I called chapters, and three situations within each chapter: a room, a trip, a routine. For every situation we prepared both a song and a photograph from our own archives. Only one of the two was ever shown, so each situation met a single first cue. I am P06, and my results are marked.
Each square is one encounter, numbered by how strongly the person felt they were living that moment again, from 0 to 6. Rows are people, columns are situations, and the colour is the cue they met. Select a square to read what came back.
P04, B3, photograph
3 of 6, re-experiencing
| Before | After | |
|---|---|---|
| Pleasantness | 75 | 50 |
| Activation | 35 | 70 |
“final internship period”
P04 B3, recorded account; the stored photo inventory said “Laptop at hotel”
Thirty seconds
Each encounter followed the same sequence. The participant rated how pleasant and how activated they felt. Then a song played, or a photograph was shown, for thirty seconds. Then they described whatever came to mind before any other question was asked.
One encounter, step by step. Run it at its real length.
Music brought back the feeling more strongly
Asked how strongly they felt they were living a moment from their past again, on a scale from 0 to 6, people gave music a mean of 4.53 and photographs 3.80. Each chapter held two cues of one kind and one of the other, so I compared the two within each chapter first, weighted the chapters equally, and then averaged across the five people. Four leaned toward music, one came out even, and nobody leaned toward photographs.
Each row is one person, and the marks show how many points higher music scored than photographs in each of their two chapters. Right of zero favours music; the figure at the top is the average of all five.
How likely the gap is to be chance
With five people, a gap this size would turn up by chance about 7 times in 100 even if music and photographs worked equally well, which is just above the conventional threshold of 5 in 100. So it is a direction, and a consistent one, four people toward music and nobody toward photographs, though not yet something five people can settle. Seed is designed to test it with enough people and enough time for that to change.
The full method, and what this figure can and can't show, is in the research.
Each bar counts how often a gap of that size appears when the music and photograph labels are shuffled at random. The shaded tails hold gaps at least as large as the real one, 0.75.
Photographs brought back the scene
When people described what came back, photographs produced a particular event or scene in ten of fifteen encounters, and music in six. Music more often returned something less bounded, a routine or the general atmosphere of the period. My guess, which five people can't settle: music may be stronger because it leaves more room for interpretation. A song can be sad or happy, while a photograph is more objective, and you can't quite reinterpret it. That sits closer to the question than I expected, at least the part of it concerned with states. Once, a song rated five out of six was placed by its listener in a different period altogether. In that case the strength of the response said nothing about its destination, so ARC asks the person before following it.
What people said came back, sorted into five kinds, for music and for photographs.
“sitting at my desk, wearing my headphones, working on the most boring work, yet feeling energised thanks to this song, evening time, working till late, trying to pass time”
P01 E02, operator note after omission from the form
5 of 6, wrong period
Why moving closer to a feeling cannot prove a return
An obvious way to measure a return is distance. If the way someone feels after a cue moves closer to how a period feels to them, perhaps they went back. On the figure below, the idea stops working almost immediately. Two different periods can sit at similar feelings, so the same movement brings a person closer to both. The references A and B are also how each period feels when remembered today. Nobody recorded how it felt at the time.
Move the After point and watch its distance to A and B. Both can shrink at once, which is why distance alone cannot show which period returned.
| Before | After | |
|---|---|---|
| To A | 0.200 | 0.125 |
| To B | 0.200 | 0.125 |
After: pleasantness 70, activation 75
Memory as an instrument?
My first version of the interface promised to read a brain and rebuild a month. To turn the experiment and the secondary literature into a usable interface, I largely set aside what I could not know without medical involvement and scans. Everything below is what remained once I removed what it could not know, and added what the encounters showed people needed.
The four rules the interface follows, each taken from what people did or asked for in the study. Each mark draws its rule as a shape.
- 1
The archive records the cue; the meaning belongs to the person.
Plays, dates and sources are stored separately from what someone says a song means, and either can be corrected without touching the other.
- 2
A strong feeling is not a destination.
When an account points somewhere unexpected, the interface asks before following. Three of five participants wanted exactly that, two wanted to return, and none wanted it automatic.
- 3
Stopping is not forgetting.
Ending this cue and never showing it again are separate controls, and both can be reversed.
- 4
The first account is never overwritten.
Later accounts are dated additions beside the original, so change over time stays visible.
| Task | Without help | With help | Not completed |
|---|---|---|---|
| Choose an intention | 5 | 0 | 0 |
| Inspect suggestion reason | 4 | 1 | 0 |
| Correct an association | 2 | 2 | 1 |
| Respond to another connection | 3 | 0 | 2 |
| Stop resurfacing | 2 | 1 | 2 |
Correction and resurfacing were the hardest, so in ARC they are controls of their own.
How the four parts connect. What a person reports flows back into the archive.
ARC
ARC brings this together. Built on the four rules, it is an interface that returns a person, on demand, to a state of nostalgia, emotion or memory they keep in a song or a photograph. The archive is my own listening from July, August and October 2023; the study examples are P04’s. Nothing you write here is sent anywhere.
Choose a context, then follow where the cue leads, or stop.
VISIONSeedSession
- Choose
- Encounter
- Report
- Next
Whatever you choose, I show you why each cue appears, and nothing moves on until you say so.
Pick a month from my archive, or one of P04’s study photographs, in the Atlas.
- Type
- Photograph
- Why this appeared
- It’s filed under Hotel room, the context you opened.
- Filed under
- Hotel room
- Source
- P04, B3
P04 filed this photograph as ‘Laptop at hotel’ and later said it brought back their final internship period. Correcting where it belongs here leaves both records as they were.
Also in this context
There’s no timer here. In the study, each cue lasted thirty seconds.
Final day
Hotel room, a photograph, P04, B3
Final day
Hotel room, a photograph, P04, B3
Final day
Hotel room, a photograph, P04, B3
- The cue
- Final day, Hotel room, a photograph, P04, B3
- What you reported
- Nothing reported
- Elsewhere
- Nothing pointed elsewhere, so ARC didn’t need to ask.
The past selves in this study can be addressed but never verified, because nobody recorded them. Seed closes that gap going forward: it keeps a period while you are living it, as a cue, the dates and an account in your own words, so a later you has something real to be compared against.
This session
- Route
- Not started
- Accounts
- Nothing kept yet.
- Moved
- Nothing moved.
It fills in the form below with ‘Hotel room’ and the photograph. Nothing is kept until you save.
private by default; nothing leaves this browser
My seeds
They stay in this browser. A downloaded file is your own copy; nothing syncs anywhere.
Opening this browser’s records…
Later accounts
Later, you can meet the same cue again and write what comes back before reading what you wrote first. Both stay, dated, side by side.
No later account yet.
What you write here stays in this browser. It isn’t part of the study.
A seed for the future self
The question can only be answered properly going forward. Seed records how a period feels while a person is living it, then pairs that state with a new sound that means nothing yet. Later the sound plays again, and the study checks whether that exact state comes back. To count, it has to appear in the silence after the sound, sit closer to its own target than to any similar state, hold across separate days, and show in at least two kinds of measurement, of brain, body or behaviour. Even then it would describe one person and one set of sounds.
Phase 1 records reference states without music.
Phase 2 pairs new, meaningless sound keys with those states.
Phase 3 fixes a model that recognises each state, trained only on the references recorded without music.
Phase 4 crosses the person’s chosen target with a concealed key:
The expected result. After the matched key, the person's state should sit nearer the target than after a rival key or an unpaired one.
VISIONSeed, April to September 2026, research, experiment and interaction design by Kunsh Sharma
Five participants; 30 encounters; 51,344 plays; about 45 sources