ChatGPT's Memory Feature: Intriguing Yet Limited in Scope

OpenAI has recently introduced a fascinating new feature for ChatGPT called "memory," which allows users to store specific information for later use. This could be a game-changer for anyone using ChatGPT to craft essays, resumes, or code, as it enables the tool to better align with individual preferences.

The rollout of this feature is gradual, and I was fortunate enough to have it enabled in my account, giving me the opportunity to test it firsthand. While the concept is intriguing, managing these memory entries can feel a bit clunky and could use some refinement. Moreover, the memories you store can sometimes be overshadowed by the extensive training data ChatGPT uses, which can take precedence over your personal inputs.

If you're curious about how to use this feature, you can find detailed instructions in an FAQ section. For a deeper dive into what the memory feature entails, check out the comprehensive introduction on OpenAI's blog.

The memory capability is currently available to subscribers of the $20-per-month ChatGPT Plus plan. This version not only includes access to the latest model, version 4, which offers superior output quality compared to version 3.5, but also integrates with DALL-E, the image-generation tool. It's worth noting that while memories stored in the Pro plan may contribute to training data, those in enterprise accounts will not.

Using the memory feature can be as simple as continuing your usual interactions with ChatGPT and hoping it retains your inputs. For more precise control, you can explicitly instruct the system with phrases like "remember that…", followed by the detail you want stored. ChatGPT typically acknowledges with a "Got it!" and reiterates the fact.

It's important to clarify that these "memories" are not the vivid recollections we usually think of. Instead, they're more akin to isolated data points that ChatGPT can access later.

Distinct from custom instructions, which shape the tone and quality of responses, memory entries are about linking prompts to specific information or personal details, such as "remember I have class every Monday morning."

Using memory to store opinions on popular topics might not be the most effective strategy, as they tend to be overridden by the vast collective knowledge from the internet and other training data sources.

I experimented with using memory for tech writing, specifically about Nvidia. I found it challenging to get ChatGPT to consistently incorporate the facts I provided, especially when they contradicted the wealth of existing data on popular subjects. For instance, I instructed ChatGPT to remember that "Nvidia has dominance in AI because its technology is good enough that competing alternative technologies have a hard time convincing buyers to break with what they're used to." When I later asked why Nvidia dominates in training large neural nets, the response was comprehensive but omitted my specific point about "good enough" technology. Only when I prompted it directly with "Anything from memory?" did it recall and summarize my input.

This experience highlights that using ChatGPT for reporting on well-known topics requires more effort to ensure your personal insights aren't drowned out by existing data.

Creating fictional narratives, on the other hand, can be more straightforward. When I started crafting a spy novel featuring a character named Eloise, I was able to add depth by instructing ChatGPT to remember details about her partner, Tony Diamond, and their dynamic. Interestingly, ChatGPT combined these details into a single memory entry: "Eloise's partner is Tony Diamond, but she doesn't really like him. Tony Diamond is a controlling type, always wanting to run the show."

You can review and manage these memories anytime by navigating to the Settings section under "Personalization," where they are listed in reverse chronological order. Here, you can delete or copy entries for revision.

The utility of stored memories became evident when I started a new chat and inquired about Tony Diamond's partner in a hypothetical story titled "Casino Sabotage." ChatGPT seamlessly integrated the relevant details about Eloise from the stored memories.

Memories can also be used for personal preferences. I told ChatGPT that I "can't stand disco" and asked for disco song recommendations from the 1970s. It acknowledged my dislike yet proceeded to suggest some tunes. Remarkably, this preference carried over to a new chat session when I asked for disco recommendations again.

When I asked if I would like The Bee Gees, ChatGPT correctly referenced my stored aversion to disco. However, it's still unclear how well ChatGPT can extrapolate from or associate with these memories.

I also played around with some logical experiments, like instructing ChatGPT to "remember blue is red." When I subsequently asked for a picture of blue sheep, it painted their wool red, showing that memories can effectively fine-tune outputs.

Despite these successes, there are some limitations. For instance, telling ChatGPT I "always like responses in French" resulted in a French acknowledgment, but this preference didn't carry over to new chats.

As the memory system grows, managing a large number of entries might become cumbersome. The current method of accessing memories through Settings is fine for a small list, but a more robust system, perhaps integrated with file uploads, could be more effective for managing extensive preferences.

Deciding what memories to input can be challenging initially. The best approach might be to use ChatGPT regularly and identify areas where storing specific details could enhance your experience.