A user asks ChatGPT to plan a trip to London. The assistant already knows his travel habits, the projects he is working on, the document format he prefers to receive and certain constraints that regularly come up in his discussions.
Until now, such continuity has been more science fiction than digital assistants. Early language models were capable of producing sophisticated responses, but remained fundamentally amnestic. Each new conversation involved reconstructing a context; the preferences, projects or constraints mentioned the day before could disappear as soon as a new session was opened.
With the evolution of its memory management system, OpenAI no longer only seeks to record information. The company is now trying to connect projects, habits, preferences and decisions to build a more coherent understanding of each user.
Behind what could be seen as a simple functional improvement, perhaps lies the next great battle of artificial intelligence: that of the user model.
Memory becomes a strategic asset
Since the appearance of ChatGPT, competition between major AI laboratories has mainly been based on the power of the models. Neural network size, training data volume, reasoning capabilities or execution speed were the main differentiating factors.
As model performance converges, the ability to accurately understand each user becomes an increasingly important competitive advantage. An assistant capable of contextualizing a request through months or even years of interactions can produce more relevant responses than a system with a theoretically more efficient model.
OpenAI’s new memory architecture now attempts to distinguish what is a one-off event from what constitutes a lasting characteristic. A professional project can be completed. A preference may change. A habit can disappear. A new constraint may appear.
This refreshability is probably more important than the memory itself, because it gradually transforms the assistant into a system capable of maintaining a coherent representation of its user rather than a simple accumulation of information.
The digital twin leaves the factory
The concept of a digital twin did not originate in conversational artificial intelligence. For several years, manufacturers have been using digital representations of engines, factories, supply chains or energy networks to simulate their behavior. These models make it possible to anticipate failures, optimize operations or test different scenarios before their implementation in the real world.
The objective has never been to perfectly reproduce reality, but to construct a sufficiently faithful representation to understand the functioning of a system and anticipate its evolution.
AI assistants gradually apply this logic to the individual. Where the industrial digital twin models physical assets, the personal digital twin seeks to represent preferences, goals, habits and constraints. The raw material no longer consists of industrial sensors but of conversations, documents, calendars, emails, searches or digital interactions.
AI no longer only models conversations
Until now, conversational assistants have been designed as query processing systems. Their role was to understand a question and then generate an answer.
With the arrival of persistent memories,The assistant no longer only learns what the user asks. It begins to learn how this user works.
What topics come up regularly? What objectives do his demands pursue? What criteria influence its decisions? What constraints structure its activity?
Each interaction gradually enriches this representation.
An investor who regularly analyzes startups may see the emergence of an assistant capable of integrating his usual evaluation criteria. A manager can instantly find the history of strategic thinking carried out over several months. A consultant can rely on a system that already knows its sectors of activity, its reporting formats and the recurring problems of its clients.
The value no longer lies only in the answer produced, but also in the prior understanding of the person asking the question.
Why AI agents need this memory
This evolution is directly linked to the emergence of autonomous agents. The goal of major players in the industry is no longer just to answer questions but to execute complete tasks on behalf of their users.
Organizing a trip, preparing for a meeting, filtering information, managing a calendar or carrying out certain administrative procedures requires a thorough understanding of the user context.
An agent incapable of understanding the preferences, habits or constraints of its user will remain limited in its action capabilities. Conversely, an agent with a detailed representation of its environment will be able to make more relevant decisions with a higher level of autonomy.
From memory to user model
To construct a truly useful representation, conversations are unlikely to be enough, and memory is only the first step.
Large technology players already have considerable sources of information. Emails reveal professional relationships. Calendars outline work priorities and habits. The documents describe current projects. Search histories reflect interests. Financial applications inform economic behavior. Connected objects sometimes document sleep, physical activity or movement.
Taken in isolation, each of these flows has a limited value, but aggregated in the same system, they make it possible to construct a much richer representation of the user.
Microsoft has privileged access to emails, meetings, documents and collaborative tools via Microsoft 365. Google controls a significant part of the personal digital environment through Gmail, Calendar, Drive, Android and Search. Apple relies on its hardware and software ecosystem. OpenAI is gradually enriching ChatGPT with new data sources and the integration of new tools.
Behind the different approaches lies the same ambition, that of building the most complete digital representation possible of their users.
Who will own our digital double?
This development raises questions that go far beyond the protection of personal data, because the issue is no longer just the raw data, but the model derived from it.
Who controls this representation? Can it be transferred from one assistant to another? How to correct an erroneous interpretation? How can we audit the mechanisms that construct this digital portrait? Who decides what constitutes a lasting characteristic or a one-time behavior?
These questions could quickly become central as assistants gain autonomy.
This perspective recalls the reflections of Stanisław Lem in Summa Technologiae, published in 1964. The Polish writer already imagined the appearance of systems capable of constructing artificial representations of the world and individuals from a massive accumulation of information. His question was not so much about the intelligence of machines as about their ability to produce models rich enough to reproduce certain human behaviors.
More than sixty years before the major models of language, Lem foresaw a question which today becomes concrete: from what moment does a digital representation cease to be a simple file of information to become an operational model of an individual?
