What We Risk When AI Systems Remember
In April 2025, while announcing improvements to ChatGPT’s memory, Sam Altman expressed his excitement about “AI systems that get to know you over your life,” promising that this would make them “extremely useful and personalized.”
This kind of personalized lifelong knowledge capacity in AI systems represents a fairly recent innovation. It involves a form of long-term memory called non-parametric memory, in which information is stored in external files rather than being embedded within the AI model itself. By default, AI systems could access information only within a limited context window, typically restricted to the current conversation. This constraint is analogous to human working memory, which can only hold a few items in active awareness at any given time.
The expansion of memory capabilities isn’t unique to OpenAI’s ChatGPT; other companies, including Anthropic and Google, have implemented it in their respective AI systems. Given that such developments are likely to transform how users interact with AI, it’s important to question whether lifelong, personalized knowledge actually enhances their usefulness. This article will examine the design of long-term memory, the risks associated with personalization, and recommendations for mitigating harm.
Related Knowledge Center: Artificial intelligence
On top of a failing business model, newsrooms must navigate the powerful — and not fully reliable — new capabilities of artificial intelligence and the deployment of large language models.
State of the business model | Copyright and intellectual property