# How to Save Travel Preferences for ChatGPT and Claude

Every AI trip-planning session starts the same way. You type three paragraphs explaining how you travel. The assistant produces an itinerary. You correct it twice. You get something decent.

Then the conversation ends, and all of it evaporates. Next trip, you type the three paragraphs again.

The fix is not a better prompt. It is writing the preferences down once, somewhere your assistant can read them every time.

## What native memory gets and what it misses

ChatGPT and Claude both carry memory now, and for travel it is genuinely half-decent. Mention a few times that you would rather walk a neighborhood than queue for a landmark, and it picks that up.

Where it stops is the exact stuff.

Your partner's shellfish allergy. The fact that your budget is flexible on hotels and not on flights. That you will not do a guided tour, ever. That your knee means under six flights of stairs or you are carrying the bag up yourself.

Those do not reliably survive as inferences, because you mentioned them once, in passing, eight months ago. And they are precisely the facts that decide whether an itinerary is usable or fiction. An assistant that forgets your taste gives you a slightly generic trip. An assistant that forgets a constraint books you into a fourth-floor walkup with a seafood restaurant downstairs.

Written preferences do not get forgotten, and you can read back exactly what your assistant is working from.

## The four things to write down

### 1. How you travel

Pace, lodging, food, transport, money.

> We travel slow. Three nights minimum per place, ideally more. Apartments over hotels, always, mostly for the kitchen. We eat one real meal out per day and cook the rest. Trains over flights whenever it is under six hours. Budget is flexible on where we stay and tight on flights.

### 2. Hard constraints

The non-negotiables. Be blunt here, it is the highest-value section in the file.

> Marta: shellfish allergy, serious. My knee: no more than about six flights of stairs, and check that lifts actually exist rather than trusting the listing. School holidays only, so July and August. No guided tours.

### 3. Places you loved, and why

The "why" is doing all the work here. A list of cities tells your assistant nothing it can reason with. A reason transfers to somewhere you have never been.

> Porto, 2024. Best trip we have taken. Walkable, no single must-see attraction so there was no pressure, and we spent the week in the same three cafes. That is the shape we want again.
>
> Barcelona, 2022. Too much. Every day had a monument in it. Came home tired.

Now an assistant can suggest Ghent or Bologna and explain why, instead of listing whatever ranks highest on a travel blog.

### 4. Where you want to go

The wishlist, with any context attached. "Japan, but not in summer, and Marta wants to see it before her parents get too old to come."

## Setting it up

The [Travel Profile template](/templates/travel-profile) is the shortcut. It creates the four folders and ships the instructions your assistant follows: search the profile before recommending anything, treat Preferences and Constraints as binding, and flag a conflict rather than quietly pushing through it.

That last one matters. Those are [folder instructions](/docs/ai-instructions), so they go to the assistant every time it opens the folder, instead of depending on you remembering to mention the knee and the allergy in the prompt. It is guidance to a model, not a hard gate, so read an itinerary before you book it. But the constraints are in front of the assistant every time rather than eight months back in a chat log, and that is most of the battle.

Connect [ChatGPT](/docs/chatgpt) or [Claude](/docs/claude) over MCP and both read the same profile. Write it once, use it in whichever one you happen to open.

## Then the planning session gets short

You open a chat and say: "We have nine days in September. Somewhere in southern Europe. Check my travel profile first."

No preamble. No three paragraphs. It already knows the pace, the allergy, the stairs, the fact that you want another Porto and not another Barcelona.

And when you come back, add two lines about how the trip went. What worked, what you would not repeat. The profile gets better every trip instead of resetting to zero.

That is the difference between an assistant that plans trips and one that plans *your* trips. Same model, both times. The only thing that changed is that you wrote it down. If you want the longer version of this setup, the [travelers use case](/use-cases/travelers) walks through it.