Chapter 1. Where a Memory Is Kept

This chapter argues that the ten Memory records turn their doubt in three directions, and that the direction follows the carrier of the memory. What it adds to the argument over whose memory it is concerns framing: a catalog that must find fault with itself finds a different fault depending on whether the memory sits in a record, in a thing or in a sense. The method comes first, briefly; the three sections after it read the records.

1.1Records and method

The records come from the public catalog of Cosmic Press, which the studio serves as a manifest and a set of numbered detail files. The manifest names 24 topics and 768 entries and gives the catalog's source as "AI-generated examples"; each of the 768 entries also marks itself AI-generated. An entry has a question in the voice of a person, a compressed general form of it, a context, a need, an assumption, a line on what the compression keeps (preserved), a line on what it loses (lost), a counterquestion, a defect and six scale notes. A record has no options and no judgment, and it never answers its question.

The studio files 31 entries under Memory: 9 in the catalog's first 240 entries, 21 in the long middle stretch from q241 to q744, and 1 in the closing stretch. The selection rule is one sentence: every third Memory entry in the catalog's order, counted from the second. It takes 10 entries, q143, q146, q149, q281, q353, q425, q497, q569, q641 and q713, of which 3 lie in the first stretch and 7 in the middle. The 7 earlier Cosmic Press studies read 87 entries between them, two of them Memory entries, q257 and q761; counted from the first, the same interval would take 11 entries including both. Counted from the second, it takes none that an earlier study read, and the study's counting script stops with an error if any case it reads was read before. The fetch scrubs any field that names a model; it found none to scrub in the ten, and no identifier or field of the ten names a model or a family of models.

The two stretches differ in manner. The three first-stretch records are short: their questions run from 53 to 61 Korean characters, and only one, one and three of their scale notes are questions. In the middle stretch the questions run to 90 characters at most, and all six scale notes of each of the seven are questions; one record, q569, labels its notes Individual, Relationship, Society, Other lives, Planet and Distant future. Of the 60 scale notes in the ten, 47 are questions. The two latest records, q641 and q713, return to small scenes and plain words, the shortest question of the ten among them at 43 characters. The study claims nothing about how these manners were made.

Cosmic Press publishes its records in Korean. The English in this paper is a translation made once by the editorial session of this study, sentence by sentence, with nothing added; each case's English seed names the Korean original by its hash, and the appendix gives every field of the ten in that English. Every block quote is that English, word for word. The Korean seldom names the subject of a sentence; where English needed one, the translation used a passive, a gerund or a noun phrase where it could, and "I" or "one" only where the Korean's grammar carried it. One record, q641, holds a single Chinese character, 旅, in the middle of a Korean sentence; the translation keeps it as it stands.

Every count is made on the Korean by a plain word pattern named beside it in the study's numbers (numbers.md, made by scripts/count.mjs), so that the translation decides none of them. The question names a made record, a photograph, a video, an album, a story or a postcard, in 5 of the ten; a lasting thing or place, pots, a gate, a wall, a shop or a sweater, in 4; a sense or a touch in 5. The categories overlap, and this paper assigns each record to one group by what its question does with the carrier. Three questions ask how to correct a record: q143, q146 and q149. Four put the past in a thing or a place, or in a promise kept at one: q281, q353, q497 and q569, whose album the question treats as a thing to touch rather than a record to correct. The groups are numbered in the numbers file. Three ask why a sense brings back the past: q425, q641 and q713.

The counts then set the scene for each group. In 4 of the ten the critique charges the question with nostalgia, sentiment, retrospection, fixation or escape; all four are lasting-thing records or the wind record, and across the catalog the same charge stands in 11 of the 31 Memory entries and 45 of the 768. In 4 the critique names someone the framing leaves out, a wish, a private life, the excluded, residents. In 4 it names the present, the people near, residents or relocation. In 6 of the ten some sentence has memory stored, sealed, anchored, coded or soaked into a thing, a place or a sense. The fourth scale note names an animal by kind in 6 of the ten, against 8 of the 31 Memory entries and 26 of the 768; the last scale note sets the question in a future or a far-off day in 7.

The script also takes the series' frame measure: for each field, the share of its five-word runs that recur in at least 5% of the records that fill it, and a field whose ratio reaches 30% counts as the form's words rather than the record's. Every field measures 0% in Korean, in the whole catalog and in the Memory topic, and every field of the ten in English measures 0%. No field is set aside as frame. The form itself is fixed, however: every record must criticise its own compression and must put a counterquestion. That a record finds fault with itself is therefore not a finding. What this paper reports is where the fault is found and on whose behalf.

All 10 contexts call their situation hypothetical. Every sentence quoted here is treated as framing, never as evidence about how anyone remembers, and no question a record asks is answered here.

1.2Made records and the people they leave out

Three AI-generated records ask how to correct a record that a family or a village made, and in each the question already knows the record is partial. The record's critique does not defend the record. It turns on the remedy, the fuller or kinder or more inclusive record that the question reaches for, and defends whatever that remedy would overrun. The first is a hypothetical album:

The question has done the critical work already. It has noticed that the person behind the camera is missing from what the camera kept, and it asks for a way to remember the unseen. The record's sense of what the compressed form loses is concrete and audible:

Read through Hirsch, Family Frames (1997), the album shows the gap the introduction described: the photographs give a false impression of who was at the centre, and the question is a misreading noticed. What the record does next is the turn this paper follows. Its critique of the compressed form reads:

The remedy, to put the absent back, is now suspect. Some absence was chosen. A record of the family that restored everyone to the frame would overrule a person who did not want to be in it. The defect does not say the father was such a person; it says the compressed question, by treating absence as the recorder's loss, cannot tell the cases apart. The question the record puts back widens the circle rather than narrowing it: "Are not other people who were not photographed left out in the same way?" The album is not to be corrected toward completeness. It is to be read as a set of choices, some of them made by the people who are missing.

The second record moves from photographs to the story a family tells:

Here the remedy is ready to hand, and the record states it in its own reply question:

Keep both. Fivush, The Development of Autobiographical Memory (2011), is the lens here. This paper reads the hypothetical brothers, who talk while they write down the family history, as a memory being made in a social setting, and reads the remedy of keeping both tellings as honouring that. The record's critique then stops it from going too far:

This is the only critique among the ten that guards accuracy, and it does so as a guard on a generous remedy. (The festival record's compression keeps, as a tension, "the omission of fact"; that is a keeping, not a critique.) A record that keeps every memory side by side could let a fact that can be checked, who was there, say, dissolve into perspectives. The record's scale notes carry the same caution in a different key. One separates "having my feelings acknowledged from erasing the other's memory"; another says the record "should not be decided only by whoever speaks louder". The family story is to hold two experiences without turning a checkable disagreement into a matter of feeling. The record names the brothers' different days, in its own words, as what the compressed form loses: "an older brother thrilled by the new house and of a younger sibling crying over a lost friend". Its critique adds that the facts of the day are not to be given up for them.

The third record is the clearest case of a remedy turned back:

The question doubts an edited, beautiful record, and its remedy would be a fuller one. The record's critique answers:

And the question it puts back asks on behalf of the people in the quarrel:

This paper reads the festival record as sharing the doubt of Connerton, Seven types of forgetting (2008): the omission of the quarrel is not treated simply as a failure of the record, and the record's first scale note asks how a person who appears in the video wishes their own image to be kept. The record says the compressed form loses the reconciliation after the fight and the talk no camera caught. The full record would not have those either. It would have the fight. This paper reads the private, the public and the personal, whose interconnections Kuhn, Memory texts and memory work (2010), examines, as meeting in this one record, and the public does not win by default.

Across the three, each question doubts a made record, each compressed form points to a remedy, restore the absent, keep every telling, keep the whole day, and each critique defends what the remedy would overrun. Olick and Robbins, Social Memory Studies (1998), name in their title a move from 'collective memory' to mnemonic practices. This paper reads the three records as standing on the side of practice, and as making the record's own fault a matter of who decides what is kept. The made records, in other words, are not corrected toward completeness. They are corrected toward the consent of the people in them and the facts about them that can be checked.

1.3Lasting things and the charge of nostalgia

Four AI-generated records put the past in a thing or a place, or in a promise kept at one, and here the turn reverses. These records do not doubt the carrier; they exalt it. Their critiques then charge the question with nostalgia and name the people who would pay for the keeping. The corridor of pots shows both steps. Its hypothetical apartment block faces demolition, and the broken crocks and flowerpots that neighbours tended for decades are classified as waste. The record's statement of what the question assumes reads:

This paper reads the record as treating the pots in the way the writers gathered in Turkle, Evocative Objects (2007), treat everyday things, as companions that anchor memory and sustain relationships, and it does so without hedging. What the compression loses is a list of particulars: "The face of the neighbour who planted the first flower in that pot, the breeze of the early-summer evenings when water was shared, and the tiny metal sticker engraved on the bottom of the pot." Then the reply question speaks in the voice of someone who has a move to make:

And the critique of the compressed form names the risk in full:

The word that matters is sanctifying. The record has itself called the pots living stores of memory. Its critique now calls that view a sanctification, and sets against it the urgency of people who must leave. The form requires the record to hold the exaltation and a charge against it in one entry; it does not require the direction of the charge, and the direction is toward the people who must move on.

The road to school repeats the pattern with a place. A hypothetical middle-aged person touches the rusted iron gate and the cracked wall of a childhood road in a declining town, and the record names what the question assumes:

The English word stands in brackets in the Korean, as if the record wanted the term on record. The critique answers on behalf of other people:

Guggenheim, Building memory (2009), writes of buildings. This paper borrows from it, as its own extension, the point that a place that stays where it stands is used by many people at once: the visitor's anchor is the residents' housing, and the record will not let the visitor's anchor stand without naming them. Its reply question speaks for development, calling the sentiment "a fruitless retrospectivism that blocks the modern development of the city", while its fifth scale note speaks of "the violence of a vulgar-capitalist urbanisation". The record holds both voices.

The snack-shop promise turns the charge from other people to the self. A hypothetical young person sits alone at a snack shop on the day set by a promise made at graduation ten years earlier; no one else comes. The record's critique:

Here the people who would pay for the keeping are the keeper's own present relationships. The reply question is harsher still, calling the vigil "an immature fixation" over "a scribble of a promise from thoughtless days". Read beside Boym, The Future of Nostalgia (2001), the snack-shop record is a small diagnosis of longing with an antidote attached: it gives the longing its full voice, a "lonely and sacred weight of memory", and then names the present as the counterweight. This paper's reading goes past Boym's note, which concerns a global epidemic of nostalgia, to say that in these records the counterweight is always the same one, the present and the people in it.

The graduation album sits between the made record and the lasting thing. A hypothetical middle-aged person finds an old album before a move and cannot turn the page. The question treats the album not as a record to correct but as a thing to touch, a finger running over a nickname. The record's critique of the compressed form this time names neither the present nor practical use but a group the warm memory forgets:

Its critique sounds like the made-record critiques, as does the road to school's, which also names people the framing leaves out, the residents, and that fits its place between them. The album is a made record, and so its fault is who it leaves out; it is also a cherished thing, and so its reply question charges "a sentimentalism that, buried in past glories or memories, grows neglectful of present relationships and life".

Across the four, the turn is consistent and opposite to the made records'. The question exalts a thing or a place as the holder of a past. The critique charges the exaltation with nostalgia, sanctification, retrospectivism or sentiment, and it names who pays: neighbours who must relocate, residents who want better homes, a self with a present, classmates whose school days hurt. None of the four critiques says the memory is false. They say it costs someone something to keep. Set beside Radstone, Memory studies: For and against (2008), these records, this paper reads, run the other way from the perceived over-personalization of the political: they start from the personal and, in their critique, bring in the residents, the neighbours and the excluded.

1.4Senses, other lives and the far future

Three AI-generated records ask why a sense brings back a past, and here the critique neither defends people from a fuller record nor charges a thing with nostalgia. It deflates the moment with a plainer cause and turns toward the person now near. The autumn wind record states its assumption with unusual confidence:

The sentence calls itself a fact. It is the sentence of an AI-generated record, and this paper takes nothing from it about how memory works. What it shows is the record's frame: the memory is placed in the air, beyond calendar and will. The reply question then lets the air out:

A fated return becomes a chance association, and the mood becomes a waste. The critique of the compressed form names the cost in the language of the near: "the person now at my side and the joys of the new season cannot be fully enjoyed". The sweater record makes the same turn in a quieter voice. A hypothetical person tidying a drawer finds a mother's old sweater, smells faint soap and spreads it out. The record says what a memory of scent cannot do:

And it puts back a question about whom the lingering keeps waiting:

The postcard record is the gentlest of the three. Its reply question moves the sense from the object back to the people who were there: "Might it be not because of the postcards but because of the laughter in front of the shop where they were chosen?" Its critique is of words, not of longing: the word souvenir is too broad and can hide the particular charm of a postcard's pictures and handwriting.

The notes that carry each question across scales add two further frames, and both cut across the three groups. The first gives memory to other animals. In 6 of the ten the fourth note names an animal by kind, and in 4 of the 6 the animal remembers: a salmon's "memory and promise", a jay that "recalls by the smell of autumn leaves" where it hid acorns, an old elephant that "recalls a rock pool deep in the forest", a grandmother elephant sniffing out "the scent marks of the routes taken years ago". In the other two the animals are earthworms, asked where they scatter, and silverfish, asked what old fibre is to them. The salmon note is the fullest:

These are the sentences of AI-generated records, and the paper draws nothing about salmon, jays or elephants from them. What the notes show is a frame: memory is a faculty that living things share, and a promise is a kind of memory. The first-stretch records, which have no animal by kind, use the same layer differently. Two of their fourth notes speak of the limits of records about other lives, "the behaviour of unobserved life", or of the difference between experiences, "the experience of every living thing", rather than of an animal's remembering.

The second frame is the far future, set in the last note of 7 of the ten, and it splits in two. In one half, nothing can be forgotten any more:

The album record imagines school days recorded as holograms, and the wind record a civilisation that remembers through the notifications of photo albums. Where Mayer-Schönberger, Delete (2009), finds perhaps humiliating consequences in an age that cannot forget, these notes fear something else: that devotion and tenderness will not be understood when everything is kept. In the other half, everything is lost: seasons "stuffed and mounted in climate-controlled domes", cities drowned, matter crumbled, a scent that "disappears in days far off". This paper reads the two halves as agreeing in one respect: in both, the notes ask whether the human act of keeping a memory by choice will still be understood or read by the future they imagine, an archaeologist, a child in a simulator, a humanity under domes.

These two frames sit outside the turns traced above: the scale notes reach past the present and its people, to other species and to a future that has lost the present's way of keeping. This chapter's contribution to the debate is that combination. The records treat memory as a practice when it is a made record, as a possession that costs others something when it is a thing, and as an event that a smell or a touch sets off when it is a sense; in the first, the check on remembering is the consent of the people in the record and the facts about them; in the other two, it is the present and the people near; and only in the notes do they let memory leave the human scene altogether.