Let's start simple. Your Wi-Fi goes out. Last time, Chat or Claude helped you troubleshoot it. You reset the adapter, the connection came back, and you asked it to remember what worked.
A week later, the Wi-Fi drops again. This time the assistant takes you through other fixes and heads toward something more complicated. You stop it. “Wait. This isn't what we did last time. Can you check?” It finds the earlier note and changes course.
Why did you have to remind it? The information was there.
Saving a note is one part of memory. Bringing it into the conversation when it matters is another. Even if the assistant had found it immediately, there's still a question: should it tell you to reset the adapter every time the Wi-Fi goes down?
Of course not. This time the router could be the problem. The old fix is worth checking, but it doesn't get to decide today's answer. That's the part of agent memory I find more interesting. How does an agent use an experience without turning it into a rule?
Remember what happened. Question why.
I'd want the record of that first Wi-Fi problem to say what we actually know. The connection dropped. We reset the adapter. The connection returned. Did we check whether the router came back at the same moment? If we didn't, we shouldn't write down “the adapter was the cause” as though we'd proved it.
That may sound finicky, but it's how bad lessons get made. Something worked once, so the agent reaches for it again. Then again. Before long, a useful clue has become a habit nobody thinks to question.
Now suppose the second outage really is the router. Keep both experiences. The first isn't wrong because the second was different. But the agent's conclusion should get more careful: the same symptom can have more than one cause. Next time, it should compare the conditions before suggesting a fix.
And I'd want to see why. What did it conclude from the first case? What did it observe this time? What remains uncertain? If it can't answer those questions, it has collected stories, not learned much from them.
A small business has its own versions of this. A booking confirmation doesn't arrive. An invoice stops syncing. Someone remembers the last fix, and everyone is tempted to try it again. The cost of a wrong guess may be higher than a few minutes without Wi-Fi.
Make the experience useful next time
I wouldn't want to call up a special memory feature every time I suspect a problem has happened before. The system should look for relevant past cases as part of its normal work. It could show a strong match, a case that might be related, and the conditions that don't line up. Then a person can decide what to try.
Several cases might even point to a pattern nobody noticed at the time. Fine—but show me the cases that support it and the ones that don't. Until we check it, it's a possibility, not a fact.
But whose cases should it be allowed to compare? A client's private history can't become somebody else's shortcut just because it might help. If access ends, the agent must stop using that case. Revising a conclusion is another decision. So is deleting the record.
I'm talking about the system around the model here: its records, evidence, and feedback. The model itself doesn't retrain after every Wi-Fi call. This is an idea we'd need to test before saying it helps agents make better decisions.
How might that work? Start with a record of each outage: what went wrong, what we tried, and what happened next. Keep that history as it happened.
Then keep the conclusion separate. After the first outage, the agent might note that resetting the adapter helped. After the second, it has a fuller picture: the router can cause the same symptom. Next time, it should check what failed before suggesting either fix.
The harder question is how the agent finds the right past case and knows whether it fits today's problem. That's what I'd explore next.
For now, when it says, “I've seen this before,” I want to hear: “Here's what happened last time. Here's what looks the same. Here's what doesn't.”
That gives us a place to start. The past helps the agent ask a better question today. That's what I mean by learning from experience.