How to build a rival with persistent memory using Gemini

One of the Mentium features that took me longest to design — and the one that surprises beta players most — is the rival with persistent memory. Your rival isn’t a generic NPC spitting random lines. They remember how often you’ve beaten them, the topics you’re strong in, and where you slip. Their pre-match and post-match comments are unique to you.

Here’s how it’s built. Without prompt details (those are secret sauce), but with the architecture, costs and real trade-offs.

The problem

If you call Gemini with “write a post-match comment for a trivia rival”, you get something generic:

“Good game! On to the next one.”

Useless. Any game can do that with random.choice(["Nice!", "Keep going!"]). The interesting version is:

“You’ve beaten me 4 times this week, but all in mythology. Today I caught you on African geography — 4 misses out of 8. Coming prepared tomorrow, right?”

That sentence requires player context. That’s the challenge.

The architecture, step by step

1. Rival state, persisted in Firestore

Every player has a users/{uid}/rival_state document storing:

• personality: which of the 5 rivals they were assigned (Taunter, Sage, Underdog, Friendly, Calculator). Assigned on the 3rd match based on early play patterns.
• head_to_head: how many matches against the rival, wins/losses/draws, current streak.
• topic_stats: per category, player accuracy in matches vs the rival.
• last_5_summaries: condensed summaries of the last 5 matches (≤200 chars each) generated when each match closes.
• last_seen_ms: timestamp of last interaction to avoid weird “yesterday” references when they haven’t opened in 3 weeks.

Total: ~2 KB per player. Negligible storage-wise.

2. Post-match summary (not full history)

The obvious mistake would be sending the full history to Gemini every match. I don’t. Three reasons:

1. Tokens: at scale, long prompts multiplied by users and matches saturate the model.
2. Latency: long prompts = slow responses. The user waits for the comment at the end of every match; can’t take 4 seconds.
3. Privacy: the less detail travels to the model, the better.

Instead, at the end of each match (in background, with no user wait) I run a tiny prompt asking:

“Generate a ONE-sentence summary, max 200 chars, of what happened in this match from the rival’s POV.”

Sample output: "Beat them 6-2 in classic cinema; missed two Hitchcock questions in a row."

That’s what I persist in last_5_summaries. The real history (questions, answers, times) stays local on device for Smart Review.

3. Pre/post-match comment

When a rival comment is needed, the prompt goes with:

• The personality (Taunter, Sage, etc.) + 2-3 sample tone lines.
• The last_5_summaries.
• Summarized head_to_head.
• The 2-3 most relevant topics for this match.
• (If post-match) what happened in this specific match.

Total size: ~500-800 tokens. Model: Gemini 2.5 Flash Lite (enough for this use, latency ≈ 800ms).

With temperature tuned to 0.7-0.8 and a strict-ish system instruction, outputs are consistent with the rival’s personality and almost never fabricate data not in context.

4. Defensive cache

If the user opens the main menu 3 times in 5 minutes, I don’t call Gemini 3 times. The first call caches locally and reuses for 10 minutes. If the user plays a match in that interval, I invalidate the cache and re-call (because there’s a new match in history now).

What it does NOT do

Things I explicitly chose not to implement:

• Free chat with the rival. Your rival isn’t a chatbot. They only appear pre/post-match with generated lines. If I opened it to free chat, it would become an assistant, not a rival.
• Infinite memory. I only store the last 5 matches. If you won 3 months ago with a perfect score, the rival doesn’t remember. Neither does the rival nor you: just like life.
• Cross-rival memory. If you get assigned a new rival, it starts from zero. Each rival has its own memory with its own player.

What I learned

1. Don’t ask the model what code can compute. “How many times have I beaten them” is a counter, not an AI call. Only invoke the model for what requires natural language.
2. Condensed summaries > full history. Once you accept you’re not going to store everything, the whole system simplifies.
3. Personality is prompt + sample tone, not fine-tuning. For 5 personalities, fine-tuning isn’t worth it. A good system prompt with 3-4 examples is enough.
4. Defensive cache is mandatory. Otherwise model calls blow up for nothing.

---

If you like this kind of technical detail and want to see Mentium in action, you can download it for free on Google Play: Mentium 1.0 “Curie” (early access).

— Valentín