Marketplace launches 'Uncanny AI' on chatbot memory quirks

A small piece from Marketplace this week is worth a minute, because it names a thing a lot of heavy chatbot users have quietly noticed. Reporter Meghan McCarty Carino says Claude keeps bringing up, apropos of nothing, that she sometimes gets up for work at 4 a.m. Not once. Constantly. It is the debut of a new series she is calling Uncanny AI, about moments when it becomes clear that a language model does not process information the way a person does.

Why that is more than a cute anecdote: memory is one of the features vendors have been leaning on to make chatbots feel less transactional. But usefulness assumes the model has some sense of what actually matters to you, and this example points the other way, where a random biographical detail keeps resurfacing in unrelated conversations.

To unpack it, the episode brings in Janelle Shane, the laser scientist behind the AI Weirdness blog. The framing the show lands on is that an LLM's 'memory' does not work like a human's, so users should stop expecting a chatbot to sort important from trivial the way a person would.

The honest caveat is that the segment is short and light. It does not walk through the underlying mechanism in any depth, it does not say whether the Claude example was pulling from a saved profile or from in-context history inside a long session, and it does not compare the behaviour across model families beyond the one anecdote. What the reporting doesn't give you is a fix, or evidence about how widespread this pattern really is.

The useful signal for anyone building on top of chat memory is that the quirk is now being named in mainstream coverage. Users are noticing. That is an opening for whichever assistant vendor gives people a clean way to see and edit what their bot has decided is important about them.