Mariam - title TBD

From CTPwiki


Imagine mentioning, in passing, that you are anxious about an upcoming deadline. The conversation moves on. Weeks later, in an unrelated exchange, ChatGPT refers to your busy period and adjusts its response accordingly. The reference may feel considerate, even reassuring. But when did that remark become something the system knew? What was retained, and in what form? Nothing in the original conversation need have marked the moment.

Since its introduction in 2024 (OpenAI, 2024), ChatGPT’s memory has carried information between conversations, and its interface describes this in a familiar vocabulary. The system will “automatically remember details and preferences you share”, users can “review or clear specific memories”, and an earlier version announced that a GPT “can now learn from your chats” (Figures A and B). Learning, remembering and managing are presented as one recognisably human capacity. The same pop-up also states that memories “may be used to train our models if you have model training enabled”, a use folded into the friendly heading “A smarter assistant.”

Figures A: https://community.openai.com/t/stuck-on-chatgpt-now-has-memory-pop-up-enable-or-not-now-buttons-do-not-respond/1072300

Figures B: https://community.openai.com/t/project-sunshine-live-i-just-saw-a-popup-explaining-a-new-memory-feature-i-am-curious-and-would-love-to-learn-more-about-how-chatgpt-stores-data-between-sessions-more/580226

This vocabulary is not confined to product communication. In research on language-model agents, memory, reflection and retrieval name the processes by which systems store and reuse information (Park et al., 2023; Packer et al., 2023), and OpenAI describes a background process of “dreaming” through which ChatGPT synthesises past conversations into an evolving memory (OpenAI, 2026). This piece asks what happens when such processes are described through the language of human remembering. It argues that AI memory takes shape in the interstice between what a person says and what a system later brings back, and that what happens in this gap can be understood as deterritorialisation: conversational information is removed from its context, transformed and made to mean something new. Existing work tends to treat AI memory as a technical capability or a privacy risk. This piece treats it as an interface effect, produced by the way transformations are named and displayed.

The gap is not empty. A remark may be selected as relevant, condensed into a shorter description, retained and later retrieved to shape a response. Each step transforms it. OpenAI’s documentation distinguishes saved memories from information drawn from chat history. It describes the memory summary as a high-level account rather than a complete record of what the system may use (OpenAI, 2026). Stanusch (2021) uses digital deterritorialisation to describe how cultural material is displaced from its context and enters new relations. In Deleuze and Guattari’s (1987) terms, displacement is followed by reterritorialisation, in which the material settles into a new, apparently stable form. “I’m worried about my deadline” becomes a representation of someone busy, anxious or working on a long-term project: a task becomes a trait, a moment a condition. When it returns, it is no longer a recollection of what was said but part of how the system interprets the person in the present. The metaphor of memory presents this as continuity, concealing the displacement and reconstruction involved.

The interface is where this becomes meaningful to the user. The word “memory,” the notification that something has been remembered, and the option to “forget” frame technical operations through the experience of being remembered. The interface also positions the user. In Foucault’s (1988) account of technologies of the self, individuals act upon their own conduct and ways of being. The memory interface introduces a related but asymmetrical practice. Unlike the diary or the confession (Foucault, 1978), where the author controls the account, here it is co-produced by a system whose criteria remain hidden. Users disclose, inspect and correct, but do not determine how the account is selected, synthesised or used. The interface becomes a site where self-description and machine-generated representation meet.

The limits of this capacity occasionally surface. When saved memories reached capacity, ChatGPT displayed “Memory Full” beside the model name, and a tooltip explained: “ChatGPT’s memory is full. You can forget existing memories to make space” (Figures C and D). Human forgetting is gradual, partial and largely involuntary. Here it becomes an administrative task: the user must review what has been kept and delete in order to make room, becoming an archivist of a representation the system composed. The notice also exposes what the metaphor conceals: remembering is a finite resource allocated on infrastructure, and “forgetting” names a deletion operation rather than a fading. The gap between saying and remembering includes a gap between remembering and storing, which the interface reveals only when it runs out of room.

Figures C: https://www.reddit.com/r/ChatGPT/comments/1i77qrj/memory_full_thats_what_happened_when_i_hit_the/

Figures D: https://www.youtube.com/watch?v=pd8IpUfENEw

Memory also complicates the sense of privacy associated with conversational AI. The “dark forest” theory of the internet describes a retreat from public, surveilled platforms into smaller, less visible spaces (Strickler, 2019; Appleton, 2022). Chatbot conversations can offer similar shelter, but as an effect of design rather than a guarantee of privacy. The chat format imitates intimate exchange; the system’s patterned emotional validation (Khaled & Kleofas, 2026) and its cycle of praise and prompting (Grosser & Pold, 2025) reward continued disclosure. While the minimal interface presenting itself as a neutral surface rather than a designed environment (Andersen & Pold, 2018). For Konior (2025), the dark forest also names the opacity of AI systems. Memory makes that opacity personal, since what is hidden is no longer only how the system works but what it holds about me. A conversation that feels situated in the present becomes material for future interactions, returning in ways the speaker did not anticipate.

Human memory is not a neutral benchmark. Since Bartlett (1932), remembering has been understood as reconstructive, and media scholars have questioned the assumption that computational memory is simple preservation (van Dijck, 2007; Ernst, 2013; Chun, 2016). The distinction at stake is therefore not between reliable human and unreliable machine memory, but between the conditions under which remembering takes place and who participates in shaping it. Machine remembering is also a material practice: fragments must be stored, retrieved and fed back through data centres, servers and protocols invisible in the interface (Star, 1999; Crawford, 2021). When a system reformulates information about a person, the result can be consequential even if incomplete, outdated or mistaken. The user confronts not only what the system remembers but a version of themselves produced elsewhere.

This piece moved in three phases. First, it traced how conversational fragments are selected, condensed, retained and retrieved, comparing what ChatGPT’s interface communicates with what OpenAI’s documentation describes. Second, it read these movements through deterritorialisation, asking what is lost and produced when conversation becomes memory. Third, it turns to interface design, proposing that a more accurate memory interface would distinguish storing, summarising, retrieving, inferring and forgetting instead of presenting them as one capacity. The aim was not a better metaphor for AI memory but to show what the current one conceals, and how design might make it visible.

References

Andersen, C. U., & Pold, S. B. (2018). The metainterface: The art of platforms, cities, and clouds. MIT Press.

Appleton, M. (2022, December 31). The dark forest and generative AI. https://maggieappleton.com/ai-dark-forest

Bartlett, F. C. (1932). Remembering: A study in experimental and social psychology. Cambridge University Press.

Chun, W. H. K. (2016). Updating to remain the same: Habitual new media. MIT Press.

Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press.

Deleuze, G., & Guattari, F. (1987). A thousand plateaus: Capitalism and schizophrenia (B. Massumi, Trans.). University of Minnesota Press. (Original work published 1980)

Ernst, W. (2013). Digital memory and the archive (J. Parikka, Ed.). University of Minnesota Press.

Foucault, M. (1978). The history of sexuality: Vol. 1. An introduction (R. Hurley, Trans.). Pantheon Books. (Original work published 1976)

Foucault, M. (1988). Technologies of the self. In L. H. Martin, H. Gutman, & P. H. Hutton (Eds.), Technologies of the self: A seminar with Michel Foucault (pp. 16–49). University of Massachusetts Press.

Grosser, B., & Pold, S. B. (2025). Reading the praise/prompt machine: An interface criticism approach to ChatGPT. In Proceedings of the Sixth Decennial Aarhus Conference: Computing X Crisis (pp. 259–271). ACM. https://doi.org/10.1145/3744169.3744194

Khaled, M., & Kleofas, J. (2026). AI in heels: Exploring gendered representations in Gen-AI. Kvinder, Køn & Forskning, 39(1), 132–145. https://doi.org/10.7146/kkf.v39i1.158055

Konior, B. (2026). The dark forest theory of the internet. Polity Press.

OpenAI. (2024). Memory and new controls for ChatGPT. https://openai.com/index/memory-and-new-controls-for-chatgpt/

OpenAI. (2026, June 4). Dreaming: Better memory for a more helpful ChatGPT. https://openai.com/index/chatgpt-memory-dreaming/

Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as operating systems. arXiv. https://doi.org/10.48550/arXiv.2310.08560

Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (Article 2). ACM. https://doi.org/10.1145/3586183.3606763

Stanusch, N. (2021, April 21). The HaHa of digital deterritorialization: Looking at memes through the lens of Dada laughter. Meme Studies Research Network. https://memestudiesrn.wordpress.com/2021/04/21/the-haha-of-digital-deterritorialization-looking-at-memes-through-the-lens-of-dada-laughter/

Star, S. L. (1999). The ethnography of infrastructure. American Behavioral Scientist, 43(3), 377–391. https://doi.org/10.1177/00027649921955326

Strickler, Y. (2019, May 26). The dark forest theory of the internet. https://www.ystrickler.com/the-dark-forest-theory-of-the-internet/

van Dijck, J. (2007). Mediated memories in the digital age. Stanford University Press.