With Personal Agent, it becomes much easier to arrange your itinerary when traveling. Just say to it “Help me arrange the trip” and it will help you arrange it clearly.
But looking back at the sentence “Help me arrange my itinerary”, I actually omit a lot of necessary information. When to leave, whether to take a plane or a high-speed train, where to book a hotel, or whether to arrive a day in advance, these are not mentioned.
In the past, when using AI tools to arrange itineraries, it was inevitable to ask several questions back and forth. But for a Personal Agent who has been with you for a long time, you don’t need to ask repeatedly about many things. Because it can find some clues from past records. For example, you like to live near the meeting location, you tend to choose high-speed rail when traveling on business, and you are used to setting aside time to prepare materials before important meetings.
However, just because we have done this in the past, does that mean we have the same needs this time? This involves the Personal Agent’s ability to understand personal intentions:Discover personal preferences from historical records, and then combine them with the current situation to determine which preferences can still be used and which things need to be reconfirmed.
Three levels of personal intention
When we communicate with other people, we often omit information that both parties know. Just like the business trip arrangements mentioned earlier, if the other party knows you well enough, he or she can know roughly what you want without explaining many things in detail.
Personal Agent is similar. The PersonalAlign study published by ACL in 2026 divided personal intentions into three levels: explicit intention, preference intention and conventional intention.
Note: The difference between explicit intention, preferred intention and conventional intention, source @PersonalAlign original paper Figure 1
explicit intentEasiest to understand. For example, if you say, “Help me book the Shinkansen to Tokyo tomorrow afternoon,” the time, destination, and mode of transportation are all clearly stated, and the Agent can search for trains based on these conditions.
preference intentionThere are just a few more unspoken demands. For example, you only say “book a ticket for me to Tokyo”, but the Agent knows from past records that you like a window seat and don’t like to leave too early. These preferences can become its reference when selecting trains.
customary intentThen go one step further. Assume that you check your schedule and organize meeting materials before each business trip. When the Agent discovers new business trip arrangements, it may think that you need to make these preparations even if you do not request it.
The three levels involve varying degrees of inference. Explicit intentions mainly rely on current instructions, and preferred intentions need to refer to personal memory. When it comes to conventional intentions, the Agent has to further consider the circumstances under which the past behavior occurred and whether a similar situation has occurred now.
Simply put,Preferences affect how you accomplish something, and routines reflect what you are used to doing in specific situations.It is a preference to take the high-speed rail, and it is a routine to check the transportation arrangements before each business trip. The former can help the Agent improve the task conditions, and the latter gives it the opportunity to discover needs that you have not yet raised.
Preferences and routines in personal memory
After understanding the three levels of intentions, let’s look at a question: How does the Agent know what you like and what you are used to doing?
The easiest way is of course to tell it proactively, such as “I don’t like to take flights too early” and “Try to book a hotel near the meeting location.” These clearly expressed preferences can be retained and used when faced with similar tasks in the future.
But most people don’t specifically compile a list of their own preferences. More clues to their preferences are actually scattered in daily conversations, schedules, application operations, and historical tasks. Suppose you have booked ten business trips in the past, eight of which took the high-speed rail. Can the Agent judge that you like taking the high-speed rail? Not necessarily. After all, the high-speed rail may be more convenient for those business trips.
The HIM-Agent proposed by PersonalAlign provides a method to discover personal preferences and practices from historical records.
Legend: HIM-Agent’s memory formation mechanism, source @PersonalAlign
It will continue to collect user history records, first aggregate activities with similar semantics and behaviors, and then extract preference memories and routine memories respectively. Take business trip booking as an example. Although the date, city, and train number are different each time, these operations all belong to transportation arrangements. The system can classify them into one category and then compare the user’s choices in different tasks.
When extracting preferences, HIM-Agent mainly considers two aspects:historyinstructionSemantic similarity, and consistency of operation trajectories when performing tasks.
For example, in multiple business trips with similar conditions, a user actively chooses high-speed rail, which can better explain his transportation preferences than a single ticket booking. But if a user only chooses high-speed rail repeatedly on the same route, it will be difficult to judge that he has the same preferences in other travel scenarios.
Therefore, when extracting preferences, it is also necessary to retain its applicable context. Compared with simply recording “Users like high-speed rail”, “Users tend to choose high-speed rail when traveling for intercity business trips with a small time gap” can obviously provide more reference.
The extraction of routines also requires attention to the state when the behavior occurs.
For example, a user compiles their weekly work report on Friday afternoons for several weeks in a row. In addition to identifying these similar operations, the system also needs to see if they occur repeatedly at similar times and work situations.
When screening routines, HIM-Agent will combine the stability of the user’s status, the frequency of behavior, and the weight of related records to determine whether a certain behavior pattern is stable enough. This provides the opportunity to extract the memory of routines such as “at the end of the work week, users organize project progress.”
Of course, this convention may not apply if the user is on vacation or the related project has ended.
Although preferences and routines both come from long-term records, the information they focus on when extracting them is different:Preferences focus on choices and actions in similar tasks, while routines focus more on the connection between repeated behaviors and specific situations.
Matching historical preferences to current situation
With personal memory, the agent can refer to past experiences in new tasks. Going back to the example at the beginning, you say to it: “Help me arrange a trip to Tokyo next week.”
The Agent first identifies the destination, time range and task goals, and then looks for relevant information from personal memory. For example, you can find past business travel preferences through semantic retrieval, use the calendar to understand next week’s work schedule, and see what conditions can be determined in advance.
Suppose the Agent finds several memories: you like to take the high-speed train, keep your hotel as close as possible to the meeting location, and are accustomed to arriving a day early when attending important meetings. But it checks the calendar and discovers that you have a meeting in Tokyo on Wednesday afternoon and a mandatory online meeting on Tuesday night.
At this time, all past habits cannot be copied. Is it still possible to arrive a day early? Can the high-speed rail meet the schedule? These need to be re-judged based on the current time and traffic conditions.
Another situation: you have always been accustomed to arriving a day early, but this time you explicitly ask to leave on Wednesday morning. Then the Agent should make arrangements according to the requirements this time, and cannot change your decision without authorization just because there is a long-term habit in the memory.
If there is a conflict between several conditions, the Agent can also sort out feasible solutions first, and then ask you which one you value more. For example, arriving early and taking less time off can’t be satisfied at the same time, so it’s up to you to decide what to prioritize.
Therefore, personal memory can help the Agent ask fewer questions, but it also needs to be combined with the current situation to determine which preferences can be inherited, which ones need to be adjusted, and which things must be asked to you again.
From vague requests to executable tasks
After the previous processing, the original sentence “Help me arrange my trip to Tokyo next week” now has a more specific meaning.
Note: Intent recognition flow chart
Agent can divide the collected information into three categories:
Confirmed conditions:I’m going to Tokyo next week and need to make travel arrangements; the calendar shows there is a meeting on Wednesday afternoon and work arrangements on Tuesday night.
Preferences inferred from personal memory:The hotel should be as close as possible to the meeting location. The transportation method should refer to the previous high-speed rail preference. Allow preparation time before important meetings.
Things that still need to be confirmed:Where to start, what your budget is, and whether you need to arrive a day early this time.
With this information, the Agent can further arrange the itinerary. For example, you can filter transportation options based on meeting time and existing schedule, then look for hotels based on your accommodation preferences, and sort out the options that meet the conditions.
However, these three types of information play different roles in subsequent tasks.Confirmed conditions can be used as task constraints, inferred preferences can be used for filtering and sorting, and undetermined matters need to be confirmed with the user as appropriate.
Actual systems can save this information through model context, structured task status, etc., and there is no unified implementation form. The focus is on retaining the source and degree of certainty of different information, and avoiding treating historical preferences as explicit requests made by the user this time.
As for subsequent execution, the Agent can first use this information to generate itinerary suggestions. If operations such as payment and meeting cancellation are involved, it is also necessary to decide whether to execute them based on user authorization and task risk.
Continuous updates of personal preferences
Another problem with personal memory is that people’s habits change. For example, in the past, I was more concerned about the price when traveling on business, but after changing jobs, I may value time more. I used to like to have meetings in the morning, but now I may want to leave the morning for other work due to schedule adjustments. If the agent still gives suggestions based on past habits, they may not meet your current needs.
So how does the Agent know that your preferences have changed? One way to do this is to learn from subsequent interactions.
PAHF (Learning Personalized Agents from Human Feedback) provides an idea that divides this process into three links:Clarify preferences before action, leverage relevant memories during execution, and update memory based on user feedback after action.
For example, the Agent recommends high-speed rail based on past records, but you tell it: “In the future, time will be the priority for business trips, and if the flight is convenient, choose the flight.” At this time, the Agent can update the original preference memory and refer to the new requirements when arranging the itinerary next time.
But if you say “I’m in a hurry this time, let’s take a plane”, the meaning is different. You just want to fly this time. There is no need for the Agent to change your long-term transportation preferences. Just remember the reasons for this choice.
Of course, the Agent may also interpret temporary choices as long-term preferences, or fail to detect changes in your habits in time. This requires constant revision in subsequent interactions, so that its understanding of you slowly catches up with the real changes.
Conclusion
To understand your intentions, Personal Agent needs to first find personal preferences and routines from long-term records, and then determine which information can be useful based on the current task and actual situation.
But people’s ideas always change. What you liked in the past may not be what you like now; what you were used to doing may not be appropriate in another situation. Therefore, even if the Agent knows a lot about your past, it may not be able to guess your thoughts correctly every time.
This requires the Agent to know when to confirm with you while referring to personal memory, and adjust its judgment based on subsequent feedback. As these abilities are gradually improved, we no longer have to explain our needs from scratch every time.
