People Change, and Agents Should Too
People change, so personal agents need memories that leave room for new preferences, uncertain motives, and the person in front of them at each conversation.
Writing and language are a medium for consciousness. When we write down an idea, we also leave behind some of our judgments, questions, and feelings from that moment. Reading it again some time later, we may reconnect with that line of thought, or we may understand the same sentence differently because of what we have experienced since.
For AI, context and memory can also be seen as a kind of memory snapshot: the information available to one inference is kept outside the model's weights, giving something from the past a chance to enter the next judgment.
But a snapshot has a problem. It records the thoughts of that moment, not the thoughts of the future. Humans don't have the same problem, because our minds have a kind of “native language” of consciousness. Human writing and language are more like a projection of consciousness onto one dimension; for AI, they are everything in that moment. The point is that an AI's “consciousness” or “thinking” does not continue uninterrupted, so all we have is that moment.
The same is true of a personal agent's memory. It can find what a user said in the past, or even the full conversation in which those words appeared. Finding it isn't the problem. But what happens after it's found?
Today, this involves retrieving and selecting memories, as well as how the model processes context. What I want to explore is a different possibility: could we give context a similar capability? Could it tell the model not only about the current state, but also about what might happen in the future and the patterns underneath it, allowing the model to achieve a kind of self-evolution through that context?
After the Weekend
In a typical personal agent scenario, a user might share many things about their life with their agent: someone they recently met, why they are hesitating over a decision, or what they hope for in the future. These conversations don't necessarily come with a specific task, but they leave behind traces of how the user sees the world.
Take the weekend.
A user often tells his agent that weekdays are exhausting, and that he just wants to spend weekends alone. Going out and seeing people takes energy; only when he has the time entirely to himself does it feel like real rest.
The agent remembers this. Later, when the user asks what he could do over the weekend, it suggests things he can do alone: watch a film, take a walk, or stay home and read. The user accepts these suggestions. After all, in that moment, the information is correct.
But some time later, the user brings up the weekend again: “Staying home doesn't seem as relaxing as it used to.”
If the agent only remembers “the user likes being alone,” it might suggest a few different films or another walking route. It hasn't forgotten the user, and it hasn't stated anything factually wrong; it is simply continuing to make suggestions based on an understanding that worked in the past.
But this time the user replies: “It's not that I don't know what to do by myself. I think I want to see some friends—I just don't really want to go to a big gathering.”
The agent guessed wrong. The user hasn't simply offered a new activity preference, either. In the past, he felt that making plans with someone would take away from his time to rest. Now, he is beginning to find that a relaxed meeting can also be restful.
So the next time they discuss the weekend, the agent suggests meeting a friend for coffee as well as taking a walk alone. The user says that sounds good—there's a friend he hasn't seen in a long time.
This is a small change. The agent hasn't upgraded its model, modified its code, or even made any arrangements for the user. It has simply responded differently, carrying a new understanding from the previous conversation.
But is this change worth keeping?
A Pretty Good Guess
It takes just one summary to go from “wants to see some friends” to “the user prefers socializing.” That summary might still be useful for another weekend, but it could quickly become another misunderstanding.
The user hasn't said he no longer likes being alone, or that he wants company every weekend. He has only discovered that, sometimes, seeing a friend doesn't get in the way of rest. If the agent compresses this change into a new preference, it is merely replacing one fixed answer with another.
So the next time the user says, “I'm too tired this week; I just want to be on my own,” the agent shouldn't rush to conclude that he has changed again, nor should it remind him, “Didn't you say you wanted to see friends?” The earlier understanding can remain, but it needs to make room for the present situation.
Later still, one of the user's friends comes to the city for just one week. The user says, “I want to see him every day this week. Help me think of places we could go.”
This is a clear request for that particular week. The agent can use it to suggest something for each day, but it doesn't need to turn it into “the user likes socializing every day.” After the friend leaves, the user might want some quiet time alone again; or the visit might make him realize he'd like to see his friends more often. Both are possible, and the agent needs more information from future conversations.
But what does it need to record?
A rule is convenient: the next time it reads “weekend,” it knows it can suggest seeing a friend. But a rule can also easily lose the reason it was created. Keeping several possible explanations seems more honest, yet comparing them again every time increases the cost of computation and judgment. Perhaps what it should retain is a method: first understand what the user wants at this moment, then decide whether to bring up an option that worked in the past.
From some of my own independent experiments, this kind of “evolution” can produce visible results quite easily. And, objectively speaking, getting an agent to change something is not particularly difficult today.
The truly difficult part is evaluation: how does it know whether this information is “correct”?
The user accepted one suggestion to meet for coffee, which means that suggestion was appealing on that occasion; it doesn't mean he wants to see a friend every week from now on. If the user doesn't reply, he may be busy, still thinking, or may not have seen it at all. Silence itself cannot be taken as approval.
If the agent asks the user to rate “would you like me to keep doing this?” after every suggestion, the conversation becomes strange as well. The user only wanted to talk about the weekend, and now has to spend time training the agent to understand him.
This brings us back to the question we started with: a memory might be absolutely accurate for the situation in which it was formed, yet by the time the next conversation takes place, it could already be full of errors.
And perhaps there is a pattern here: the more accurate a compressed memory is at the time, the less accurate it may become in the future.
This problem also appears in the HorizonBench study. The researchers constructed 360 simulated users and up to six months of conversations for each. The users mention changes in their lives, but do not directly tell the agent, “My preferences have changed.” When researchers later ask about the users' current preferences, models often give the users' old answers. (HorizonBench, 2026)
Evolution Without an Endpoint
In this situation, what we are actually facing is a kind of logical paradox: on one hand, we want to keep improving models and tools so that they can perform sufficiently accurate semantic compression at the time and produce something useful for the user's next conversation, becoming “better and better at understanding the user.” On the other hand, we have to hope that the user doesn't change too much before that next conversation. If the user goes from wanting to spend weekends alone to wanting to socialize, for example, the previous memory has completely “expired.” We can only hope the model realizes this quickly and corrects it at runtime.
Of course, another path is to turn the agent into a riddler: don't produce precise memories or descriptions in the first place; instead, choose a few broad descriptions common across the user population. If I've never been accurate, I can never become inaccurate. I'm not joking. Most personality or psychological tests, and even many internet products, don't try to depict you with absolute precision. Instead, they try to place you into a group that has already been designed and analyzed. Although this can't guarantee that you will always fit an extremely well-matched profile, at least you won't have a bad experience because you've been put into an inaccurate one.
But from my perspective, today's context and memory give us an entirely new way of thinking. They are no longer generic information, but unique information created by each individual user. This uniqueness can be understood in two ways: first, everyone's life and trajectory are different, making it hard for two people to have exactly the same experiences; second, at the data level, even when describing the same public event, people's tone, punctuation, and even emoji are unlikely to be exactly the same. If the information is unique, trying to take data with all these differences and fit it into a generic profile may create even more problems.
Let's take another look at the word “evolution.” If we set aside specific scenarios and discuss the concept itself—for example, Darwin's framework of evolution—we can find some very interesting things.
For example, there is no absolute right or wrong in evolution itself. Birds don't have wings because, many years ago, all the individuals made a correct decision that “our descendants must have wings,” and then started growing them. Rather, over the course of history, offspring in a population develop different heritable variations. In a particular environment, these differences can affect individuals' chances of survival and reproduction, and the associated traits may gradually change over many generations. Individuals that left no descendants didn't evolve “wrong”; they simply didn't pass their traits on in the environment of their time.
Evolution doesn't have a fixed direction, either. If you're interested, you'll find that even for a body form like the crab's, similar crab-like forms have evolved multiple times in different evolutionary lineages throughout known history—and some lineages later lost some of their typical crab-like features. I find this quite interesting. (Wolfe et al., 2021)
On the first day, you have a phone case, decide you don't need it, and throw it away. On the second day, you accidentally knock your phone. On the third day, worried about knocking it again, you buy a new case. Looking at the result, throwing it away on the first day seems entirely unreasonable. But for living things, changes in form don't happen because they know what will happen in the future and then decide whether to keep a trait.
So, returning to our original question, my answer is that evolution in the context dimension works the same way for AI. This evolution has no right or wrong, no direction, and there cannot be a unified evaluation standard at the individual level.
In practical technical terms, this evolution is essentially a semantic compression of the human-agent interactions at the time. But the goal of compression is no longer to produce a general summary. Instead, it is to have the agent speculate about the user's underlying motivations and inner thoughts, and make some predictions about what might happen in the future. When the user tells the agent that he wants to socialize on the weekend, ideally the result of the previous round of evolution should already have included this possibility—for example, that wanting to spend weekends alone might be a short-term state rather than a long-term pursuit, and so on.
Overall, self-evolution in the context dimension is a dynamic process: its results are dynamic, its direction is dynamic, and its evaluation is dynamic. Only the user is static, serving as the environment that selects among the agent's evolutionary results, retaining the part he wants at that moment.
Beyond the Snapshot
Back to that first weekend.
The agent may notice that the user's thoughts have changed, but it doesn't necessarily know why. Is it because he has made new friends, because something has changed at work, or because he has developed a new understanding of what “rest” means? All of these are possible, but none can be determined from a single conversation.
The agent can ask, or it can leave the question open for now. It doesn't need to rush to explain the user to himself.
As for the future, the agent can recognize that the user's thoughts will continue to change, and pay attention to those changes in future conversations. But “what might happen” should not become “what the user will inevitably become.” He might see his friends more often, or he might need more time alone again at some point; those are questions for his future life and his own choices to answer.
Writing leaves traces of the past, but the person reading those words changes. A personal agent needs to remember what the user once thought, and also be capable of noticing that the user is changing. It can try to understand why, but it also needs to know it might have guessed wrong.
The same is true for AirJelly. This little blue jellyfish on your screen will quietly observe and accompany you, help you plan all your tasks, and come to know the future you.
When an understanding from the past no longer fits the present, a personal agent should be able to let go of what it once learned and get to know the person in front of it again.
By Alan Yuan ·