You’ve done this. You open ChatGPT, start a conversation, get something useful, close the tab. Two days later you come back — and it’s like talking to someone who’s never met you. You re-explain your project. You re-establish your preferences. You re-describe the context. Again.
We’ve built machines that can write poetry and pass the bar exam, but they can’t remember what you told them five minutes ago.
That’s not a feature. That’s a failure.
A developer who spent three months building an alternative called Tongue put it bluntly: the biggest issue with ChatGPT and Claude isn’t that they’re not smart enough. It’s that you have to “chase particular threads” so the model has the right context. You’re not using an assistant — you’re managing a filing system for an amnesiac genius.
Think about what a real assistant does. A human assistant learns your habits, remembers your preferences, tracks your ongoing projects, and builds on every previous conversation. You don’t hand them a briefing document every morning. They already know.
Intelligence without memory isn’t an assistant. It’s a very smart stranger you keep meeting at a party.
Here’s where most people get it wrong. The conversation in AI right now is obsessed with capability — bigger models, better benchmarks, more parameters. Who’s winning the reasoning race? Which model writes better code? But the real bottleneck isn’t raw horsepower. It’s context persistence.
The difference between “an AI that can answer anything” and “an AI that actually helps you” isn’t intelligence. It’s whether the system maintains a rich, evolving model of you across every interaction. Do you prefer short answers or detailed ones? What projects are you working on? What did you decide last Tuesday? What matters to you?
Right now, you carry all of that in your head and manually inject it into every new conversation. You’re the context layer. You’re doing the work the machine should be doing.
The real unlock in AI isn’t making the model smarter. It’s inverting the entire interaction: instead of you managing threads, the assistant manages your timeline.
This is what makes the Tongue approach interesting — not because it’s technically groundbreaking, but because it identifies the actual pain. You should be able to text your AI the way you text a friend: casually, across topics, without preambles, without context-setting. The system handles the rest. It remembers. It connects. It builds.
The tension here is real. Texting is the lowest-friction interface we have — that’s why iMessage killed email for personal communication. But making texting work for AI demands something deeply un-simple underneath: a continuous, personalized memory that tracks who you are and what you need across days, weeks, months. The surface has to be effortless. The engine has to be relentless.
Every time you re-explain yourself to an AI, you’re paying a tax on your own time to subsidize the machine’s amnesia.
Most people don’t realize how much this costs them. It’s not just the thirty seconds of re-typing context. It’s the cognitive switching cost. It’s the friction that makes you not bother. It’s the question you didn’t ask because setting it up felt like too much work. How many useful interactions never happened because re-establishing the baseline was exhausting?
The current model — threads, sessions, context windows — is a design that serves the machine, not the user. It says: “Here’s a clean slate, now brief me.” A real assistant says: “I’ve got this. What do you need?”
The future of AI isn’t a smarter chatbot. It’s a system that knows you well enough that you stop thinking about the interface entirely.
We’re not there yet. But the builders who understand that memory — not intelligence — is the actual moat are the ones who’ll get there first. The rest are still optimizing for benchmarks nobody cares about.
FAQ
Q: Isn't this just RAG with extra steps? What's actually new here?
A: RAG retrieves documents. This is about the assistant maintaining an evolving model of YOU — your preferences, your ongoing projects, your decision history — and applying it automatically across every conversation. The difference between a search engine and a chief of staff is that the chief of staff already knows what you need before you ask.
Q: So what should I actually do differently today?
A: Stop treating each AI conversation as a one-off. If you're starting from scratch every time, you're wasting your own time. Look for tools that persist context across sessions — or at minimum, maintain a living document of your preferences and paste it in. The tooling will catch up, but the habit of demanding memory starts now.
Q: Isn't persistent AI memory a privacy nightmare waiting to happen?
A: Yes, and that's exactly why whoever solves this with real data ownership and transparency will win everything. The privacy problem isn't a reason to abandon persistent context — it's the moat. The company that makes you trust them with your full timeline will own the category. Everyone else will be a commodity API call.