The most expensive part of using AI is re-explaining yourself

Everyone has built the same workaround

The thing that struck us across a few dozen user interviews was not what people disliked. It was what they had built. One person had spent a long stretch assembling a personal knowledge base in a note-taking app, wired into their AI, purely so it would know things about them. Several maintain per-project instruction files that they tune by hand. Others keep a document of standing context they paste into new conversations. Different tools, different stacks, identical purpose: not having to say it again.

Nobody counts this as work

When people estimate what AI saved them, they count the output. They do not count the evening spent building the knowledge base, or the ten minutes at the start of each session getting the model up to speed, or the fact that they abandoned a promising thread because re-establishing context was more effort than doing the thing themselves. That labor is invisible because it does not look like a task. It looks like setup, and setup feels like your own fault for not being organized enough.

The New Chat button is the problem in one button

Look at what that button actually offers. A clean slate, which is another way of saying that everything you established is gone and it is on you to rebuild it. For a one-off question that is exactly right. For anything that comes back next week, it means the relationship resets to zero every time, forever. You are not working with something that knows you. You are working with something that meets you.

You cannot own a job from a cold start

This is where it connects to everything else we have been writing about. A job, as opposed to a task, is an outcome someone stays answerable for across many runs. Answerable means remembering what happened last time, noticing what changed, and knowing that a thing you flagged three weeks ago was resolved. None of that is available to something that begins each run knowing nothing. Continuity is not a nice feature layered on top of ownership. Without it, ownership is not possible in the first place.

Continuity is more specific than memory

Memory is the word everyone uses and it is too vague to be useful. From what users actually reacted to, continuity has to satisfy at least three conditions, and the third is the one most often missed.

• It has to survive a change of surface. The single most enthusiastic moment in our interviews came from someone who set an agent up in a private message, brought it into a group channel braced to explain everything again, and did not have to.
• It has to be scoped rather than global. A design student told us the thing he valued most was that separate projects did not bleed into each other. An assistant that remembers everything about everything is its own kind of unusable.
• It has to be correctable by you. If the thing it remembers about your preferences is wrong and you cannot open it, read it, and fix it, then continuity has just made a bad assumption permanent.

What to look for instead of a workaround

If you already maintain a context document or a knowledge base to feed your AI, you have diagnosed the problem correctly and solved it with your own labor. The thing worth looking for is a system where that memory is the product's job rather than yours: scoped to the project you are in, carried across whatever surface you move to, and open enough that you can read what it believes and correct it. Anything meant to own recurring work has to stop starting over, and you should not be the one holding it together.