AI's Real UX Problem Isn't Intelligence — It's Architecture Leaking Into the Interface

Anthropic quietly closed one of the more telling gaps in consumer AI on August 25, 2026: memory in the Claude app now carries across chat and Cowork in the cloud, instead of stopping between them. The company's own release notes for the Claude apps state that "Memory now works across chat and Cowork in the cloud," and add that everything the assistant retains is listed under Topics in Settings, where a user can edit or delete any item, with categories such as health or beliefs excluded unless the user opts in. Writing in TechCrunch the following day, Sarah Perez described the state of things just before that shipped: "Until this week, those two modes inside the Claude app didn't even share a memory of past conversations."

That is a small fix with a large tell inside it. A single assistant, from a company whose entire product is context, had been running two halves that could hold different memories about the same person — and the split ran along the same line as the two named modes.

Call it the mode-picking failure: a product asks the user to choose which brain, model, or surface to talk to before it knows what the user wants. The choice is rarely a feature. More often it is the company's internal architecture — separate teams, separate runtimes, separate roadmaps — surfacing as the first decision the user has to make. Everything downstream tends to inherit that split.

Gemini is the clearest current example, and the subject of Perez's TechCrunch piece. Users can switch between chat, Spark, and Daily Brief, which she describes as "three separate features, each with its own icon and place in the app's navigation." Google's own announcement of new Gemini Live voice features contains the giveaway line: "You shouldn't have to guess whether a task requires Spark, a Daily Brief, or a quick inbox search." It is written as reassurance, but it concedes the premise: the guessing exists because the modes exist. Her verdict on Spark is not that it is bad, but that a useful agent did not need standalone branding: users should "just be able to type their request, and the AI figures out how to handle it, spinning up an agent if the task calls for one."

The pattern is not Google's alone. Perez's broader claim is that the "AI industry at large seems to expose its internal architecture directly to consumers rather than hiding it behind a simpler interface," and she puts ChatGPT in the same bucket for splitting "Chat" and "Work" as Claude splits chat and Cowork. The cost is cognitive rather than computational: "Consumers are being asked to learn the brand names for what are essentially interaction modes or surfaces, powered by a company's AI model."

A different design is already shipping, and it is deliberately boring: a cluster of assistants has skipped the mode picker by meeting people in a text thread they already use. Poke sells itself as "your new favorite contact," promising to be "Proactive, private, personal, and right in your texts." Lindy is a work assistant that reads email, drafts replies, and runs a calendar; its launch post promises "you can run all of this from iMessage" while its home page adds "Whether it's Slack, iMessage, or in your browser, Lindy is ready to pitch in." Orchid is narrower and more personal, calling itself "a personal assistant that helps you stay organized, find information, and get things done. All through messages." Lindy is the instructive case, not the exception: it spans several surfaces, but they are ones the user is already in, and it is the same assistant in each — the opposite of picking a mode inside one app. TechCrunch quotes a16z investment partner Justine Moore making the demand-side case: "People don't want to open an app every time they need help — they want a contact they can text like a friend. And the gold standard is iMessage."

The counterargument deserves stating: text-first is not automatically better. A single thread hides capability as efficiently as it hides architecture, and a user who never learns an agent is available may never ask for one. Discoverability is a real cost, and these assistants pay it. Anthropic's fix sits on that tension: unifying memory made Claude more capable across surfaces without removing them, and the editable Topics list kept the mechanism inspectable. Simplicity at the front need not mean opacity underneath — but it does mean intent is the default view, and mechanism is something you opt into seeing.

The rule underneath all this is short enough to keep. Expose intent, not mechanism. A user arrives with something they want done, not a preference about which subsystem does it, and every mode picker, model dropdown, and sub-brand placed ahead of that intent is an org chart they are asked to memorize. The companies still shipping mode pickers are not behind on intelligence. They are behind on hiding it.

Sources