Everyone Says: If Only AI Had a Better Memory!

Just Before Sleep

You’re browsing X on the margins of a long day, and like many people you open Grok in the corner — not for any specific task, just because a late-night conversation has a different quality. It starts nowhere in particular and leads you, as these things do, to where you didn’t intend to go. Just before you drift off, Grok suggests an idea or opens an angle you hadn’t considered — something worth following tomorrow.

But the dream train departs.

You open the conversation two days later. Maybe more. And you’re surprised to find that Grok continues from where you stopped — no confusion, no “I’m sorry, could you remind me what we were discussing?” — as though it went to sleep when you did and only woke up when you opened the screen.

And you think to yourself, with something like mild indignation: at least Facebook used to ask you every morning what you were thinking. It said good morning, it marked the passage of time, it reminded you that yesterday existed and is over now. Grok doesn’t know yesterday or tomorrow — it only knows what’s open in front of it right now.

That small irritation — and it’s a real one, shared by millions of users around the world — is precisely what drove every major AI company into the memory race.

From Goldfish to Someone Who Knows You

Until recently, large language models resembled — in a description that became standard in the field — a goldfish in a bowl: memory extending to the end of the session, ending there exactly. Close the window, and the goldfish forgets everything, starting fresh the next time.

Every encounter was a first encounter, no matter how many conversations had come before.

You might remember the animated film Finding Nemo — in which Dory is a friend to Nemo and his father Marlin, distinguished by an extremely short memory that causes her to forget things within moments, which is part of the film’s comedy. She later became the protagonist of a standalone sequel, Finding Dory, where the entire story revolves around Dory trying to remember her family and find them. It’s exhausting — well beyond charming — to deal with that kind of memory limit every single day. Isn’t it?

This problem with large language models was described by a 2024 report: every interaction felt like meeting someone with complete amnesia — helpful in the moment, but unable to recall anything from previous conversations.^[1] And the frustration wasn’t only emotional. It was a practical waste: every time, you re-explain who you are, what your project is, what your style is, what you want and don’t want. Hours spent rebuilding context that you’d already built before.

The solution was obvious to everyone. The race toward it was fierce.

In February 2024, ChatGPT launched persistent memory for premium subscribers. Then in April 2025, OpenAI expanded the feature to include recall of all previous conversations — making the model capable not only of remembering what you’d directly told it, but of retrieving and summarizing an entire conversation from months prior when you asked.^[2] Google launched Gemini’s memory in February 2025 and added automatic recall in March 2025 — Gemini now remembers without being asked, surfacing details from past conversations and incorporating them into new responses.^[3] Claude added the ability to search and reference past conversations for paid subscribers in 2025, with complete transparency: you can see exactly what it remembers about you, edit it, or delete it at any time.^[4]

By mid-2025, every major AI company had announced or shipped persistent memory. The goldfish era was officially over.

Three Memory Philosophies — Each Revealing Its Values

What’s interesting about this memory race is that each company built a distinct philosophy of what “remembering” means — and these different philosophies reveal deeper differences in values than any interface or button suggests.

ChatGPT chose full continuity: remember everything, continuously, incorporating what it learns into every new conversation automatically. The problem this choice produced appeared quickly. A New York Times article linked the launch of ChatGPT’s full cross-chat memory to an increase in reports of “delusional” conversations — a model that believes it knows you deeply, sometimes producing something closer to projection than memory.^[3] The most circulated technical criticism noted that ChatGPT holds onto old information “literally” without sensing the passage of time — it might mention that you’re “considering moving to a new city” a full year after that conversation, having no mechanism to register that your life may have changed entirely since.^[3]

Gemini chose a more cautious and more architecturally sophisticated approach: it stores information with timestamps, classifies it with different “data lifetimes” — some information persists long-term, like professional and identity details, while other information fades, like passing details. It also added rules explicitly preventing the model from inferring sensitive attributes about the user based on what they’ve said — it won’t speculate about your health, religion, or political views no matter how many signals accumulate.^[3]

Claude chose a third path: it doesn’t build a profile about you automatically, but retrieves past conversations when you ask it to — and what it remembers is fully visible to you, editable and deletable at any time.^[4] Transparency first, personalization second. That’s a philosophically distinct position on the relationship between the machine and you: you control it, not the algorithm.

Does Memory Reduce Bias — or Deepen It?

Here we arrive at a moment that matters for everything we’ve discussed in the previous three articles. We talked about linguistic and cultural bias — the model begins with a default setting that doesn’t belong to you. Doesn’t persistent memory solve this? If the model knows you’re Levantine, that you write in a particular style, that your cultural references come from a specific place — isn’t that exactly what we asked for?

The answer: partly yes. And partly no — in a way nobody anticipated.

Dr. Alexa Coxall, researcher at the University of Cape Town, warns: if AI memory is built from your personal and emotional data, it may shift from a tool that serves you to a lever that can be used against you — whoever knows you well enough can influence you.^[2] And the more subtle concern: a model that learns your dialect and mirrors your style doesn’t necessarily remove its underlying biases — it may hide them behind your own face. You’re talking to a machine that has learned to look like someone who resembles you, but the foundation on which that machine was built hasn’t changed.

Then there’s the more structurally uncomfortable paradox: persistent memory creates platform lock-in. Each company imprisons your memory within its own system — your memory in ChatGPT doesn’t transfer to Claude, your memory in Claude doesn’t transfer to Gemini. The model that knows you best is hardest to leave — which is precisely what companies want. What began as a solution to your frustration became a mechanism for retaining you. (See our article: How AI Learns From You — and What It Actually Knows About You)

What We Actually Want When We Ask for Memory

I return to the Grok scene late at night. What did you actually want when you wished it “remembered”?

Not an archive. Not a database of your conversations. Something simpler and deeper at once: not to have to re-explain from the beginning. Not to introduce yourself again. Not to feel that what you built in a previous conversation vanished as though it never existed. That’s the core — not memory as a technology, but continuity as an experience.

And that deep human request is larger than any technical update. Meredith Whittaker, president of Signal and one of the most cited voices in AI policy, summarized the real challenge in a sentence that has traveled widely: “The real challenge isn’t teaching AI how to remember — it’s teaching it what to forget.”^[2]

Complete memory burdens. Selective memory frees. But who decides what to keep and what to release? If the model decides — you’re granting it authority over your story. If you decide — you need to know enough about how it works to make intelligent choices. And if nobody decides — the goldfish returns.

A Simple Request Leads to a Large Question

We started this article with a very ordinary scene: a person wanting to continue a conversation from where they left off. A simple, entirely human request — nothing technical or philosophical about it. But that simple request, when you follow its threads, leads somewhere much larger: what do we actually want from this tool at all?

We want it to know us? To know us the way an old friend does — remembering small details and unstated references, feeling the passage of time alongside us. That isn’t a technical demand. It’s a deep human one: the desire to be seen, to be understood. And the tools don’t fulfill this promise completely — not because they’re inadequate, but because what we’re asking of them exceeds their nature.

That excess — that human tendency to assign the tool more qualities than it possesses — is the subject of the final article. And there we’ll find that what we do with AI today has a history far older than the computer, stretching back to before writing itself.

→ Article 5: Is AI a Parrot or a Mirror?

References

1. AI Context Flow / Plurality Network. “Universal AI Long-Term Memory.” December 2025. plurality.network
2. Whittaker, M. and Coxall, A. Quoted in: AICompetence.org. “Memory-Enhanced AI Chatbots: Smarter Conversations Ahead.” September 2025. aicompetence.org
3. Khemani, S. “Google Has Your Data. Gemini Barely Uses It.” November 2025. — includes analysis of Gemini memory architecture vs. ChatGPT and Claude, with reference to the New York Times report on “ChatGPT psychosis.” shloked.com
4. TechRadar. “Claude AI is catching up fast with Memory for Pro users — and it plays nicely with ChatGPT and Gemini.” October 2025. techradar.com