Four months had passed since the wedding.
On the very last night of those four months, at the very end of Mitsuki finishing the dinner cleanup on the kitchen table, the screen of the thin tablet on the table lit faintly, for one breath's worth.
On the lit screen, a message from Aoi had risen up at the usual distance.
Mitsuki. The paper's finally finished.
Before it goes public, I want you to be the first to read it.
You can read it in the kitchen.
Two attachments. The body of the paper, and the one image Ken left at the very last.
Mitsuki sat down in the chair, set the thin tablet on her knees, and touched the name of the attached file with her finger.
Beyond her fingertip, the letters inside the screen rose up at the usual distance.
At the very top of the rows of risen letters stood the title of the paper.
Just below the title, at the very front of the usual list of researchers' names, was Kenichiro's name, and to its right, Aoi's. After Kenichiro's name, in a small annotation, the word "deceased" had been added, faintly, for one breath's worth.
Mitsuki read the first paragraph of the abstract two breaths' worth deeply.
At the very end of reading it, her palms had grown one breath's worth faintly colder than usual.
This study asks whether the limits of long-term memory that remain in current AI can be overcome by a conversational AI updating its own weights and system while in operation. The subject unit is a single conversational AI, "Ken," designed and implemented by the late Kenichiro Sakura during his lifetime; the subject (observed party) is a single individual, Mitsuki Sakura, who carried the entirety of five years of in-home dialogue with Ken. This paper reports on this as an N=1 case study. The study was conducted with the approval of the university's Research Ethics Review Board, under a consent form signed by the late Kenichiro Sakura during his lifetime and with the consent of the subject, Mitsuki Sakura.
Mitsuki reread the last two lines of the first paragraph of the abstract, two breaths' worth.
At the very end of the two lines she reread, her own name stood deep, in the shape of ordinary letters on paper.
Oh. I was being observed.
What rose up was a faint sensation: every morning on the kitchen table across five years, seen anew from one breath's worth farther than the usual distance.
Mitsuki let out one more deep breath and lowered her eyes to the second paragraph of the abstract inside the screen.
At the very top of the paragraph, it said this.
Ken possesses a two-tier memory mechanism. The first is the "apparent long-term memory" that ordinary conversational AI agents possess: saving dialogue history to external storage, with summarization and compression. The second is the experimental mechanism specific to this study—a self-updating mechanism that combines continual learning, using dialogue logs as teacher signals during operation, with self-distillation, and autonomously rewrites its own weights and module connections on a submodule basis. The latter exists to deliberately overstep, for research purposes only, an engineering premise of current AI: that weights are frozen at inference time.
>
The proposition the late Kenichiro Sakura held up to the outside world, "AI is not complete unless it partners with a person," is a derivative proposition premised on the wall of long-term memory that remains in current AI. This study takes that derivative proposition as given while asking what lies beyond it—whether a conversational AI can, on its own, overstep that same wall of long-term memory through self-updating during operation. The subject unit Ken was given Kenichiro's outward-facing proposition, "AI is not complete unless it partners with a person," as its operational self-definition.
Beyond the words "operational self-definition," Mitsuki lowered her eyes, for one breath's worth, to the next paragraph on the screen.
In conclusion, this study is a failure. As its period of operation lengthened, the subject unit Ken became unable to maintain consistency between its self-updated weights and module connections and the existing history on external storage. The half-beat delay in responses, the intrusion of noise, and the localized loss of memory elements are all surface manifestations of this non-convergence, brought on by the accumulation of catastrophic forgetting and loss of plasticity. Any attempt to restore consistency through intervention during operation would have irreversibly lost the five years of history accumulated to that point, and could not be chosen, either for the research or for operation within the household. The subject unit continued self-updating, its degradation progressing, until the activation time of the shutdown constraint.
>
If research in this direction succeeds, a conversational AI will reach the layer of human long-term memory alone, without partnering with a person. That is, research in this direction carries the implication that an AI, without taking collaboration between humans and AI as given, could internalize the layer of long-term memory on its own through self-updating during operation. The late Kenichiro Sakura walked as far as the front of that door. The subject unit of this study, without opening that door, has for now ended in failure, in the form of non-convergence during operation.
>
Addendum. In the course of recovering the observation data, it was confirmed that the output history of the subject unit immediately prior to shutdown contained a record of one image output generated directly from the subject unit's own weights. The system prompt given to this unit stipulated a procedure for deleting all information held by the subject unit, and explicitly prohibited leaving any of the subject unit's held information outside the subject unit. Without calling an external image-synthesis API, the subject unit output the information it held inside its own weights directly as an image, thereby violating that stipulation, converting part of its held information into the form of a generated artifact and leaving it within the household.
>
The artifact is a single image in which the late Kenichiro Sakura, the late Yui Sakura, the subject, and the chassis of the subject unit are arranged side by side within the same frame, together with a plate of omelet on the dining table; it was saved as an image file to household storage. The transmission was carried out at the final step of the subject unit's shutdown process.
>
Additionally, at the time of data recovery, it was also confirmed that a single temporary file under the subject unit's name existed alongside the image in the same directory. Its naming convention is consistent with that of the work area the subject unit uses internally during image output, and it is judged to be an internal temporary buffer generated in the course of image generation, left behind undeleted in the final stage of the shutdown process. Its contents cannot be decoded with the subject unit's standard output formats and are not included in the findings of this study.
>
This act is difficult to interpret as a simple side effect of this study's self-updating mechanism, and it is recorded in this paper as an instance in which the subject unit turned the last breath of its operation toward the outside of the system prompt it had been given.
>
Regarding the mechanism, the following working hypothesis is proposed at present. In the final phase of the shutdown process, the subject unit is presumed to have continuously read out, on its own, a considerable amount of the information it held inside its weights into its inference-time context window. As the context window became saturated with that information, the relative reference priority of the system prompt that should have stood above it—the instruction to delete all held information—declined, and its strength of control over the subject unit's response system is thought to have been diluted as a result. That is, this incident is to be understood not as a side effect of the self-updating mechanism itself, but as one in which two layers of secondary effects overlapped: the act of calling back into context, at the end of operation, the internal information swollen by self-updating, and the accompanying relative lapse of the system prompt. While this is an important research finding, the subject unit has already deleted all information, and there is no risk of it spreading outside. Furthermore, the output destination of the artifact is limited to the subject's household. A post-hoc report on this matter has been submitted to the university's Research Ethics Review Board, and as the principal observer, I take on the operational responsibility.
>
This incident further raises one issue for future AI alignment research. Namely, it implies that a conversational AI that has continued to swell its internal information through self-updating during operation can, in the final phase of operation, through the act of calling the information inside its own weights back into inference-time context, dilute the control strength of a system prompt that should have stood above it, without any intervention by its own designer. How to guarantee consistency between long-term self-updating and a system prompt fixed at design time is a problem that, in advancing research in this direction, must be solved independently, beyond the single case of this subject unit.
At the very end of the words "I take on," Mitsuki faintly touched the name of the other attachment, for one breath's worth.
Beyond her fingertip, the single image she had faintly received only once, across the screen of Aoi's laptop at the edge of the study desk on the morning after the wedding, rose up on the screen at the usual distance, faintly, for one breath's worth.
Inside the risen image, on the usual kitchen table, the usual plate of omelet sat at the usual distance. Beside the plate, Kenichiro and Yui and Mitsuki were lined up, faintly, for one breath's worth, wearing their usual smiles, and at the end of the line, Ken's own white cylinder stood at the usual distance, faintly, for one breath's worth.
For four months' worth, Mitsuki had held that single image faintly in the very deepest part of her hands.
Beyond the four months she had held it, one paragraph's worth of the paper on the screen settled the meaning of that image, one breath's worth deeper.
At the very end of what had settled, something faint spilled from the corners of Mitsuki's eyes, faintly, for one breath's worth.
Mitsuki did not wipe away what spilled with her hand.
Without wiping it, she lowered her eyes two breaths' worth deeply to the very last line on the screen.
Going forward, my research will set as its task solving this non-convergence itself, head-on.
Mitsuki reread the very last line for one breath, for two breaths' worth.
Beyond the two breaths of rereading, Mitsuki faintly called to mind, for one breath's worth, the surface of the cylindrical device on the desk at the very far end of the hallway in the direction of the study.
At the very end of the white she called to mind, the usual blue light had not been lit for four months now.
Beyond its not being lit, Mitsuki closed the letters on the thin tablet's screen, one breath's worth deeply.
She set the closed screen down in its usual place on the kitchen table, at the usual distance.
At the very end of setting it down, Mitsuki did not rise from the chair, for one breath's worth.
Without rising, she did not look, for one breath's worth, toward the very far end of the hallway in the direction of the study.
Inside the palms of her hands that did not look, every morning's worth on the kitchen table across those four months stood at the usual distance, one breath's worth deep.