After the third request for shorter answers, an AI assistant ought to have some idea of how much detail a person wants. The same applies to a project discussed across several chats. Nobody wants to keep pasting the background just to ask what to do next.
Saving a few details would make those conversations less repetitive. Deciding which details deserve to stay is where things get awkward.
But people tell assistants all sorts of things they wouldn’t put in a personal profile. They ask questions for friends, look into purchases they never make, and discuss problems that pass. An assistant with a memory needs to make sense of that mess without treating every conversation as a lasting description of its user.
Some Details Are Worth Keeping
A preference for short answers is fairly easy to understand. Someone who always wants prices in euros could set that once too, rather than add the instruction to every question.
Planning work over several weeks gives memory more to do. Take a house renovation. The owner may have supplied the measurements, ruled out a few ideas and settled on how much to spend. When they come back to discuss the cupboards, they’d reasonably expect to pick up where they left off.
Even here, there are different kinds of information. The dimensions may stay the same. The budget might change. A suggestion discussed on Monday may have been rejected by Friday. Keeping everything without recognizing those differences could leave the assistant working from the wrong version of the plan.
There’s also a difference between using information during a conversation and carrying it into future ones. A person may be perfectly happy to explain a situation so the assistant can answer today’s question. That isn’t necessarily an invitation to bring it up next month.
A Question Isn’t Always a Preference
Sometimes a question says very little about the person asking it. Someone finding a birthday present may spend an afternoon comparing fishing equipment without ever intending to go fishing. A parent looking into a course for their child isn’t necessarily interested in taking it themselves.
Someone reading about the casinos operated by Estolio Limited on Online Casino Groups might ask an AI assistant to clear up a few payment terms. If they come back to the comparison later, remembering which pages they’ve already checked would save them some time.
They probably wouldn’t appreciate casino suggestions popping up in unrelated chats, though. Asking about something once doesn’t mean they want it treated as a lasting interest.
Several questions about the same subject can look like interest, even when the person is simply trying to finish a job. Once the job is done, there may be nothing more to carry forward.
Asking before saving a broad assumption would give the user a chance to explain. “Keep this with the comparison” is a different instruction from “remember that this interests me.” The assistant should have room for both.
People Change Their Minds
A request for the cheapest hotel can make sense for a single night near an airport. It gives the assistant little reason to assume that price will matter more than everything else on every future trip. If that original request becomes the default, later suggestions could be wrong for reasons the user doesn’t immediately recognize.
People also move, change jobs, finish courses and abandon plans. Having an assistant keep referring to a previous workplace would get irritating quite quickly.
Someone shouldn’t need to argue with it about an old job or a finished project. A correction ought to change the saved detail too. Otherwise, the same misunderstanding could turn up again the following day.
When a saved detail came from a guess, that should be apparent in the reply. “You asked about this last time” gives someone something they can recognize and correct. “You like this” makes a much larger claim.
Memory becomes more useful when its origins are visible. A short note showing that a preference came from a particular conversation could help settle a misunderstanding before it affects another answer.
A Shared Screen Makes Things More Complicated
Picture an assistant open on a kitchen tablet. More than one person may use it, but the account on screen still belongs to one of them. A question from somebody else could then be answered using details from that account holder’s earlier conversations.
That might only produce an unsuitable dinner suggestion. It could also bring up something the account holder had expected to keep to themselves. People sharing a screen need to know whose saved information the assistant is drawing on.
The account name should be easy to spot, with separate profiles available where shared use is expected. There should also be an easy way to begin a conversation without drawing on remembered details, particularly when showing the assistant to somebody else.
Being able to read the memory matters here too. A user shouldn’t have to wait for an embarrassing reply to discover what the assistant has been carrying between conversations.
Forgetting Should Be an Ordinary Request
Finding a saved preference shouldn’t involve searching through weeks of chats. There ought to be a place to read what the assistant remembers, with enough explanation to tell where each detail came from. A wrong entry could then be corrected directly, instead of relying on another conversation to put things right.
The word ‘delete’ deserves some care as well. Stopping the assistant from using a remembered preference and removing stored conversation records are different things. Someone clicking a removal button needs to know which of those things it changes.
A task could also have an end date. Holiday plans don’t usually need to keep influencing answers once everyone has returned home. The renovation discussion may need to last longer, but the person doing the work is better placed to judge that than the assistant.
There will be people who prefer to start fresh every time. They shouldn’t have to accept memory just to use the rest of the service.
For someone who does use memory, the test is fairly ordinary. Can they return to useful work without repeating everything, then move on when it’s finished? An assistant that keeps dragging an old question into the next conversation hasn’t saved them much effort. It has given them another thing to sort out.
