# How Kult Uses Gemini for Game Help Without Relying on Model Memory

Kult's gaming companion separates patch-aware knowledge, voice and second-screen UI. Here is how to test the same design before scaling it.

By Paras Madan | 2026-11-21


If someone asks “How do I get past this boss?” in the middle of a game, knowing which boss they mean is only half the problem. The other half is knowing which version of the game they are playing, what gear they have, and how much of the answer they can listen to without putting the controller down.

In a [November 20 architecture write-up](https://cloud.google.com/blog/products/gaming/kult-built-a-gaming-companion-with-gemini-that-actually-knows-your-game), Kult founder Omid Ashtari describes an alpha that uses Gemini for spoken game help and pairs it with a second-screen feed. Kult reports coverage of more than 100 titles. That is a claim about breadth, not a benchmark for accuracy. The useful part is the separation between learning about a game, finding the right answer in a session, and presenting it to the player.

## Put the current patch in the answer path

Kult says it builds structured knowledge bases per game and updates them separately from the model. That gives the conversational model current material to reason over instead of asking it to recall game mechanics from training.

For a prototype, I would start with one game and a few mechanics that have changed between versions. Put the game version, platform if relevant, and source date next to each piece of evidence. Supply the player's current context separately. That makes it possible to distinguish a perfectly good answer for the wrong patch from a failure to identify what the player is asking about.

Include an unknown-version case. If the missing detail changes the advice, the correct response might be a quick clarification rather than a plausible guess.

## Route by decision, not just by how long the prompt is

Kult describes using a lighter Flash tier for repetitive background work, Gemini Flash for conversational turns and larger models for heavier work. It also uses a decision model for some routing choices and escalates uncertain cases. The post does not publish comparable per-turn cost or latency figures.

I would give a router three decisions to make: answer from verified context, retrieve more evidence, or ask the player a question. Log which path it picks and whether the routing decision was warranted. A quick answer grounded in the wrong guide can be more costly than a slightly slower lookup.

Measure the time until the first useful spoken instruction separately from the time until the complete visual answer. If you only measure the first token, you can optimize for a fast greeting that leaves the player waiting for the actual help.

## Make the spoken answer and the screen do different jobs

In [Ashtari's Google Cloud interview](https://cloud.google.com/transform/in-the-clouds-omid-ashtari-kult-gg-interview-ai-companion-for-gamers), he describes voice alongside a feed that can add things such as video or an interactive widget. Those surfaces make sense together because the player doesn't need to hear every detail a guide can show.

Try giving the player one next action by voice. Put the map, the alternate route and the supporting source on the second screen. Review the two outputs together: a good spoken answer paired with a visual card for a different game version is still a failed turn.

[A2UI's catalog documentation](https://a2ui.org/concepts/catalogs/) describes a contract between the agent and renderer covering components and functions. A small catalog for a game companion could start with a source card, a map image and a numbered checklist. Validate the requested components and their data, and fall back to a plain answer if a card cannot be rendered correctly.

## Test the limits before adding more titles

Kult describes using a model registry and checking replacements against answers written by people who know the games. For a first evaluation, I would add cases with contradictory sources, a player's vague reference, and a clear request to avoid spoilers. Score the answer for applicable version, source quality and usefulness while playing.

Don't infer current real-time screen understanding from the companion's session awareness. The November write-up describes game vision using Google's Live models as future work. The alpha architecture is a useful reference without treating that roadmap as a shipped capability.

The first milestone I would trust is a small set of repeatable cases where the assistant gives the right next step for the right game state. Adding a hundred more games won't fix an answer path that can't do that reliably for one.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical decisions and product behavior through clear writing and practical storytelling.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
