The Double Edge of ChatGPT's New Memory Function

Evolving Capacities; Lingering Questions

Greetings, Dear Readers,

A few weeks ago, upon logging into my ChatGPT 4 account, I was greeted with a notification about a new feature: “memory” across chats. This piqued my curiosity immediately:

Could this be the next step toward AGI—an AI form of human working memory?

Or was it simply another example of the tech industry repurposing cognitive terminology for technological uses?

In search of clarity, I reached out to my Substack network, and numerous individuals, including the insightful Daniel Nest, came to my aid. Daniel gently reminded me that he had covered this update in his newsletter back in February. It seems I had missed that while engrossed in other pursuits.

Memory and Higher-Order Reasoning

During that time in February, I was deeply absorbed in a series of fascinating podcasts by Daniel Bashir at The Gradient. One particular episode set me on an exploratory path into Daniel Kahneman’s seminal work, Thinking, Fast and Slow, for the first time. I know, dear readers, don’t hold this gap in my reading against me for long. I hope to rectify this state of affairs in this article. In this pivotal podcast, Bashir interviews computer scientist Subbarao Kambhampati, discussing topics like planning, reasoning, and interpretability in the age of large language models (LLMs).

The Gradient

Subbarao Kambhampati: Planning, Reasoning, and Interpretability in the Age of LLMs

Together, they outline a prevailing view within the machine learning community: current LLMs are limited in performing higher-order reasoning tasks, partly due to their lack of working memory. Essentially, while our most sophisticated prompting techniques aim to replicate working memory, they notably fall short of fully compensating for this capacity, which is foundational to a wide array of cognitive, emotional, and intellectual processes.

Kahneman: System 1 vs. System 2 Thinking

Throughout their discussion, Bashir and Kambhampati frequently referenced Kahneman’s concepts of “System 1” and “System 2” thinking. While I initially grasped the gist of this distinction from their dialogue, my curiosity led me to delve into Kahneman’s own writings to fully appreciate the resonances of his ideas in their original context. Kahneman’s text is a wild ride in case you haven’t read it. Kahneman prides himself on creating scenarios that his readers can replicate to test out the validity of his ideas.

Source: Neurofied

Back to the theory: System 1 thinking, as Kahneman articulates, operates automatically and quickly, with little or no effort and no sense of voluntary control. It encompasses what we might consider our instinctual responses to stimuli, those immediate, often subconscious reactions that guide much of our daily decision-making. In contrast, System 2 thinking is deliberate, effortful, and orderly. It’s the mode we engage when faced with complex problems or decisions that require focus and analytical thought.

In the lively online discussions I had with others about Kahnman’s distinction and its application to AI, I found that the reception was mixed. Guy Wilson, one of my favorite subscribers, recommended Patrick House's 19 Ways of Looking at Consciousness, as an alternative way of thinking about thinking. My own intellectual journey had previously led me to Bruno Latour's 14 modes of existence, and we cannot forget Wallace Stevens' “13 Ways of Looking at a Blackbird.” Despite the temptation to expand modes or perspectives via multiplicity, I keep returning to the System 1 and System 2 distinction. In particular, I find it thinking compelling to the extent that it highlights working memory as a dynamic target or limit for future AI systems.

What Is Working Memory?

Working memory or long-term memory is a fundamental cognitive function, akin to a mental notepad that temporarily holds information necessary for tasks such as reasoning, comprehension, and learning. This capability is crucial for advanced human activities—it underpins abstract thinking, strategic planning, and complex problem-solving, which are central to innovation and sophisticated reasoning. Moreover, working memory is essential for language processing, aiding in the construction and understanding of complex sentences, and it plays a critical role in social interactions, enabling us to consider others' perspectives and respond appropriately in real-time.

Source: “Episodic, Procedural, and Semantic Memory”

More than just a functional tool, working memory is a core component of human consciousness that significantly enriches our interactions with the world and one another. This cognitive capacity is integral to System 2 thinking, which involves managing and manipulating multiple pieces of information to address complex issues or make informed decisions, embodying deliberate, analytical thought. In contrast, System 1 thinking operates on a more automatic, intuitive basis, largely circumventing the need for the short-term storage functions provided by working memory.

Is GPT Memory Function “Working Memory”?

What role does the memory function play in ChatGPT's operations, particularly when compared to the concept of working memory in human cognition? According to Tiernan Ray in a recent exploration of ChatGPT’s memory capabilities, the function strives to emulate the human facility of working memory by retaining information from ongoing interactions. This feature enables the AI to leverage previous exchanges within a session to produce more contextually coherent and relevant responses. As Ray observes, “The memory function in ChatGPT is like a fine-tuning procedure,” yet he also notes the limitations: “Getting the results you want, however, can be frustrating at times.”

Similarly to how human working memory supports intricate cognitive tasks by handling multiple information streams, ChatGPT’s memory aims to augment interaction quality by preserving continuity. Nevertheless, unlike human working memory, which is dynamic and capable of sophisticated adjustments and processing, ChatGPT’s memory function is considerably more static and confined. It predominantly recalls stored data rather than dynamically manipulating or reassessing it in response to new inputs.

Ray poignantly highlights the shortcomings of this system, stating, "The management of the memory entries is primitive and needs more development.” This emphasizes a profound disparity in depth and adaptability between artificial and human cognitive mechanisms, illustrating that while ChatGPT's memory function marks a step towards mimicking human memory processes, it still falls short of the flexible and integrative capacity of the human mind.

Higher-Order Reasoning vs. Our Privacy

In response to my inquiry regarding ChatGPT's memory function, Rachel Harrisdirected me to a poignant commentary by Walter Haydock in the Deploy Securely Substack, which introduces a critical dimension to the discussion of AI memory: the implications for security, privacy, and legality. Haydock articulates a delicate paradox: while the advancement of AI towards higher-order reasoning necessitates the integration of working memory, this progression requires access to increasingly extensive segments of our data. This raises pivotal questions:

1. Are we prepared to sacrifice our privacy for the sake of AI's enhanced capabilities?
2. Can the potential gains in workflow efficiency ever justify a corresponding increase in our vulnerability regarding digital autonomy?

As I reflect on these concerns, I am about to conclude with Haydock’s enlightening post. What I find particularly invaluable in his analysis is twofold: firstly, his guidance on how to disable ChatGPT's memory function—a critical option for users at this crossroads—and secondly, his approach of informing readers without dictating a specific course of action. I invite your perspectives on this complex issue. How do you perceive OpenAI's initial foray into crafting working memory for AI? Please share your thoughts and potential responses in the comments below.

Nick Potkalitsky, Ph.D.

Mr. Haydock’s Definitive Post on ChatGPT Memory:

“RIP disable chat history (2023-2024). As of yesterday morning, OpenAI will retain all ChatGPT (except Team and Enterprise) prompts indefinitely. My question:

1?? Stop using ChatGPT

Some have suggested that exit is the best choice here, to make a point to OpenAI. I admire the principled nature of these suggestions, but it’s a bridge too far for me (and probably most people).

2?? Upgrade to Team or Enterprise so you can still disable chat history

I’ve done the former for StackAware, but that is $30/month/user (2 user minimum), which is a little pricey. Enterprise is $108,000/year ($60/month/user, 150 user minimum). That’s way out of reach for most companies.

3?? Use the mobile app

As of today, when using the latest version you could still disable chat history. This might be a legacy feature that will be eliminated with the next update, though.

4?? Use Temporary Chats if and when they become available

Shortly before making this change, OpenAI published a support article describing a feature called “Temporary Chat.” This appears to reproduce the functionality of disabling chat history, but with a major caveat:

“We are rolling out to a small portion of ChatGPT free and Plus users this week to learn how useful it is. We will share plans for a broader roll out soon.”

5?? Use the API

This still has a 30 day retention period. Using it will be more challenging for non-technical users, and maintaining conversation history requires a few more steps.

It also changes the billing model from “all you can eat” for a fixed fee to a metered approach.

6?? Use the Playground

This is a little more user friendly and allows for continuous conversations, including with fine-tuned models and assistants.

I also got OpenAI to confirm that the Playground follows the API retention period (30 days).

?? Bottom line: this is a major change (for the worse, in my opinion) of OpenAI’s privacy and security posture. Adapt accordingly.

Why do you think they did this?”