Hermes learns travel preferences from a debrief

Christopher Mburu built TravelOps with Hermes Agent's own skills, search, scheduling and persistent memory. The demonstrated traveler profiles and trip dates were seeded fixtures. The planning and memory updates were real agent actions, according to his account, rather than evidence that a traveler took the suggested flights.

In the first request, Hermes loaded the planning skill, made three web searches and recommended a Dubai itinerary using stored preferences. It included an Emirates flight, a window seat, vegetarian food and a hotel choice, then stopped at pending approval. A separate visa question read the fixture's stored expiry date.

Mburu then supplied a post-trip complaint about a six-hour layover. The agent repeatedly queried memory, including failed lookups, before writing a new preference and a dated lesson. A subsequent request applied the change without being reminded of it and prioritized a hotel with a previously stored Wi-Fi rating.

There is a useful inconsistency to preserve. The first plan was already described as nonstop, while the debrief complained about a layover. The memory file said a maximum of three hours and the later recommendation said nonstop only. This demonstrates a preference-update loop, with fixture contradictions requiring human interpretation.

Prompt
Original prompt

Source: dev.to · Christopher Mburu