Hypomnemata and the New Industrialization of Memory in GenAI

Researcher: James E. Dobson

In 2025, large commercial GenAI chatbot vendors quietly introduced a new “memory” feature. “Memory,” in this context, refers to the inclusion of selected summaries of prior interactions alongside the current prompt in the model’s context window, producing the appearance of persistent state in a fundamentally stateless architecture. The contents and procedures used to extract memory fragments remain opaque to end-users, making the distinction between in-context and in-weights learning essential for understanding how LLMs operate. The memory feature is now standard among chatbot applications and now underlies applications like Claude Code.
A key aspect of contemporary LLMs is that all inputsare treated as text. Because LLMs are autoregressive, the entire context—system prompts, conversational history, stored memories—enters the model as a stream of textual input. We might helpfully cast memory-enabled GenAI applications as self-recording hypomnemata. I borrow this term from Michel Foucault, who describes them as note collections into which “one entered quotations, fragments of works, examples, and actions to which one had been witness or of which one had read the account, reflections or reasons which one had heard or which
had come to mind.” These are industrialized objects—industrialized because, as 

Bernard Stiegler argues, there has been a separation of readers and writers into consumers and producers without parallel access to the objects of inscription. In my case studies of open systems using memory features, these AI hypomnemata compress and archive information about multiple sources with different potential audiences. GenAI is bound by conditions that make it a dubious source of recall. While it may aid imaginative thinking and problem solving, it has many differences from prior recording and memory technologies. The scale, complexity, and opacity of almost all GenAI platforms have served to concentrate control and have made the industrial model of memory they represent both practically and conceptually problematic. The automatic compression of prior interactions and selective insertion of these summaries into the context window as one of many pretexts fundamentally alters interactions with these technologies. This leads to the most consequential effect of the GenAI moment: a misapprehension of the scene of writing itself and the technologies that undergird interactions with it.

About the Researcher:

James E. Dobsonis Associate Professorof English and Creative Writing at Dartmouth College, where he has been teaching in the humanities since 2012 and researching computational methods since 2003. His interests and publications span a wide range of topics, from accessing computational and data resources on large, distributed computing networks to autobiographical self-representation in American literature. Most recently, he has focused on computer vision, computational hermeneutics, and the history of machine learning and artificial intelligence. He is the author of three books:Modernity and Autobiography in Nineteenth-Century America(Palgrave, 2017),Critical Digital Humanities(University of Illinois Press, 2019), andThe Birth of Computer Vision(University of Minnesota Press, 2023). He is also co-author, with Rena J. Mosteirin, of the critical code study and poetic interpretation of the Apollo 11 Guidance Computer titledMoonbit(punctum books, 2019), andPerceptron(punctum books, 2025), a creative and critical account of the perceptron—one of the earliest and most successful machine learning devices—and its inventor, Frank Rosenblatt.