How to personalize an AI agent from the first message
Personalization is a loop: read what you know before you answer, use only what bears on the task, and save what you learn.
Configure · · 4 min read
In short
- To personalize an agent, read what you know about the user before the first message and give the model the parts that bear on the task.
- Search before any choice that depends on a preference, and save a lasting fact the moment the user states one.
- Let users see, correct and delete what the agent knows. Personalization people cannot inspect feels like surveillance.
- The first message is the hardest, because a new product has nothing stored yet. That is where outside context helps.
To personalize an AI agent, give the model what you know about the user before it writes its first reply, limit that to what bears on the task at hand, and save new lasting facts as the user states them. It is a loop of reading, using and writing. Most of the work is deciding what belongs in each step, and where the knowledge comes from on day one.
We build Configure, infrastructure to connect agents, apps and chats. The steps below apply whether or not you use it, and we point out where it fits.
What does a personalized agent actually do differently?
Ask two agents for a dinner recipe. One returns a popular chicken dish. The other knows you are vegetarian, cook for two, and have twenty minutes on weeknights, and it answers with that in mind without being told. The model is the same. The second one simply had four facts in front of it.
Personalization used to be a differentiator. Now that every agent asks for inbox access, it is simply expected. What sets a product apart is how well it uses what it knows.
Step 1: What should the agent know?
Decide this before you store anything. A useful test: would this change an answer next month?
| Keep | Do not keep |
|---|---|
| Lasting facts: role, location, household, tools they use | What they asked five minutes ago. That is conversation state |
| Stated preferences: tone, length, formats, things to avoid | Guesses about them drawn from one request |
| Ongoing work: the project, the trip, the goal | Passwords, payment details, government IDs |
| Decisions already made, with the reason | Anything they asked you to forget |
The second column matters as much as the first. An agent that stores "is a React developer" because someone asked one React question will be confidently wrong later.
Step 2: Where does that knowledge come from?
- What they tell your agent. The most reliable source, and the slowest to build up.
- What they do in your product. Choices, edits and rejections say a lot. It takes time to accumulate.
- What they already told other agents. Usually the richest source and the one a new product cannot reach by default. With the user's permission, a context layer can bring it in.
- Apps they connect. Calendar, mail and documents, for tasks that need live data.
The first two only help returning users. For a first session you need the third or fourth, or the agent starts from nothing. We cover that gap in The cold start problem in AI apps.
Step 3: How do you put it in front of the model?
Two moves cover most cases.
Read before the first message. At the start of a session, load a short overview of the user and place it in the system prompt or first turn: who they are, how they like answers, what they are working on. Keep it to what fits in a paragraph or two.
Search before a preference-sensitive choice. When the agent is about to pick something the user may have an opinion on, such as a framework, a restaurant or a writing tone, search what you know for that topic first. A profile can hold hundreds of facts. The model wants the five that bear on this decision.
Dumping everything into the prompt is the common mistake. It costs tokens, buries the relevant facts, and makes the agent bring up things that have nothing to do with the request. An agent that mentions your marathon training while you are fixing a bug is showing off, and users notice.
Step 4: How do you keep it current?
Save a lasting fact the moment the user states one, one fact at a time, in plain words. Before saving, check whether you already hold it or an older version of it. "Moved to Lisbon" should replace "lives in Berlin", and both should not sit side by side.
Treat corrections as the highest-value signal you get. If a user says "no, shorter", that is a preference, and the agent should not need telling twice.
Step 5: How do you keep it from feeling creepy?
- Show what you know. Give users a page that lists what the agent holds about them and where each item came from.
- Let them change it. Edit and delete, with the change taking effect on the next read.
- Use it quietly. Apply a preference without announcing it. "Since you are vegetarian and live in Lisbon and have two children" is unnerving. A vegetarian recipe is just helpful.
- Ask consent per source. Bringing context in from another product should be the user's explicit choice, for your product specifically.
Personalization that people cannot inspect feels like surveillance, even when it is accurate.
How does this look with Configure?
Configure gives your agent the read, search and save steps over a profile the user owns. The user connects the agents and apps they already use on a hosted page and approves your product. Your server then reads their approved context before the first message, searches it when a choice depends on a preference, and saves what your agent learns back to the same profile, in your agent's own namespace. When their context changes elsewhere, a webhook tells you to read it again.
The personalization guide shows what to build from that context: a first project, suggestions under the composer, and empty states that are already about the user. The quickstart covers the integration itself.
What is the quickest test of whether it works?
Open your product as a brand new user who has connected their context, and send one ordinary request. If the first reply contains something true and specific that you never typed into your product, personalization is working. If it reads like the reply anyone would get, go back to step 3.
Add a connect button to your product
The quickstart takes one page, end to end: a chip in your app, Configure's hosted page, and your agent reading the user's context on the first turn.
Read the quickstartNext: AI onboarding without the questionnaire