I prefer an AI that finds the right context to one that drags my entire history into every answer. More memory also means more old decisions available to appear in the wrong place. I want to know why a piece of information belongs to the current request, especially when the subject changes.
Anthropic's April 2026 research on personal guidance makes me examine that product choice. Writing in October 2026, I bring a later reference from working with agents. In June, I objected to passing the context of every job to other agents. In September, I expected them to find necessary information before returning questions.
In practice, I was asking for two compatible things: initiative in research and judgment about the material. The task did not need to become a package containing everything that happened earlier. My recorded experience concerns work; applying that concern to a personal adviser is a reflection on conversation design. The interface needs to let the subject actually change.
Imagine a deadline constraint supplied for a specific project. That example is hypothetical, but the problem is easy to understand: carried forward without its original situation, the information can resemble a permanent preference of mine. The next answer uses a premise I do not recognize. Then I have to discover where it came from.
When an answer uses another conversation, I want the relationship visible. Retrieving appropriate information helps; correcting an unsuitable association needs to be simple too. A large history without a way to revise it leaves me arguing with old conclusions while trying to formulate a new question. That is one meeting I am happy to skip.
Access to a collection of files also provides no reason to use all of them. The objective should guide the search. If an answer is in a relevant, authorized reference, I expect the agent to read it. If my choice about the conversation's scope is missing, old text should not fill that gap for convenience. Available information needs to stay attached to the situation in which it was supplied.
When assessing this kind of assistant, I will require control over the references it uses. I will look for ways to delimit context and correct an association alongside examining what it remembers. A useful answer needs a visible relationship with the current question. History can help, but I am still the person choosing the subject now.