Skip to content
FIG. 00 · ENGINE READOUT
01 WORLD STATE02 RULES03 SOCIAL GRAPH04 MEMORY05 INTENTION

Worlds that remember.

NPCs powered by a local LLM, inside an engine that stays in control.

Contact us
PLAYABLE PROOF OF CONCEPT
LLAMA 3.2 3B · LOCAL · RTX 4050 6 GB
SCROLL
01 / Why

You’ve already forgotten the name of the last NPC you met.

We are players who love stories. NPCs stayed forgettable because they don’t react naturally. We want to bring them back to life.

Classic NPC · the same line, on a loop
×1“Lovely day, isn’t it?”
×2“Lovely day, isn’t it?”
Halcya NPC · the line follows your actions
DAY 1 · FIRST VISITNew around here? Bread is two coins.
DAY 4 · HELPED MARGOTMargot told me about you. Take it, it’s on the house.
DAY 9 · STOLE A LOAFYou again. Pay first, then we’ll talk.
02 / How

The LLM speaks. The engine decides.

Follow a single action, end to end. The engine chooses an intention. The LLM only receives that intention and the useful state: it supplies the words, nothing more.

SOCIAL GRAPH · VERTPRÉEVENTPROPAGATIONOPINIONS UPDATED
BERTRAND −0.80MARGOT −0.09ÉDOUARD −0.48THEFT SEENintention → local LLM“You again. Put that back...”
PROTOTYPE NPCSNPCS IN DEVELOPMENTILLUSTRATIVE EXAMPLE
01 · THE PLAYER ACTS

You take a loaf from a market stall without paying.

Bertrand, on his rounds, sees you. The engine records an event, with one witness. Nobody else knows.

SEEN player → takes(loaf)WITNESS Bertrand
02 · THE ENGINE UPDATES THE WORLD

No global reputation: local opinions that travel.

Bertrand changes his mind about you. People close to him hear about it next, more faintly with each hop.

RELATION trust 0.30 → 0.10PROPAGATES Bertrand → Édouard −0.48 → Margot −0.09 · 2 hops (path set for the demo)
03 · THE ENGINE CHOOSES

One intention, chosen by readable rules.

refuser0.31
alerter0.22
parler_froidement0.47
RULE garde.03: theft witnessed => warning
04 · THE LLM RECEIVES THE INTENTION

It cannot change the decision. It picks the words.

Input: the intention, the useful state and Bertrand’s memory. The model runs on the player’s machine, next to the game.

LLM Llama 3.2 3B · local · reply starts ~0.6 s
05 · THE LINE

“You again. Put that back, and then we’ll talk.”

JSON valid · no invented action
06 · THE ENGINE APPLIES

Bertrand takes the loaf back.

Tomorrow, Margot will greet you more coldly. She wasn’t there, but Bertrand spoke to her.

ACTION engine: remove(loaf)
STEP 01 / 06SCROLL, OR PICK A STEP
02 / How · the decision log
“If you can’t explain a behavior by reading the logs, the system is too opaque.”

Every line leaves a readable trace, step by step. A decision can be replayed, challenged, corrected.

Runs locallyOllama · llama.cpp

No data is sent to a server. The model runs on the player’s machine.

halcya / decision.logSAMPLE TRACE · BERTRAND
t+0.000SEENplayer → takes(loaf) · witness: Bertrand
t+0.002RELATIONBertrand → player · trust 0.30 → 0.10
t+0.002RELATIONBertrand → player · anger 0.10 → 0.71
t+0.003PROPAGATESBertrand → Édouard −0.48 → Margot −0.09 · 2 hops (path set for the demo)
t+0.003RULEgarde.03: theft witnessed => warning
t+0.004CANDIDATESrefuser 0.31 · alerter 0.22 · parler_froidement 0.47
t+0.004INTENTIONparler_froidement
t+0.004LLMintention + useful state → Llama 3.2 3B, local
t+0.600START“You again…” · shown as it is written
t+2.610LINE“You again. Put that back, and then we'll talk.”
t+2.611JSONvalid · 1 line · no invented action
t+2.612ACTIONengine: Bertrand takes the loaf back
$

PROPAGATION (DEMO)Δ = opinion × trust × 0.5 × (1 + trust tendency, clamped 0.5 to 1.5) × 0.6^(hops − 1)

WHITE ENGINEORANGE LLMLILAC · METADATAILLUSTRATIVE EXAMPLE
03 / What we built

A playable proof of concept, today.

Playing loads YouTube.

DEMO 01The guard answers the player (in French)

Playing loads YouTube.

DEMO 02Exchange between two NPCs (in French)

Playing loads YouTube.

DEMO 03Édouard answers the player (in French)

The village of Vertpré

Kingdom of Aldoria · 3 NPCs · 1 social graph

Bertrand

NPC 01
ROLE
Village guard for thirty years. Investigating the thefts at the market.
PERSONALITY
Weary, fair, gruff. He sighs before answering difficult questions.
WHAT HE KNOWS
The market thefts, the bailiff’s visit in three days, the dangers of the region.
WHAT HE DOESN’T KNOW
Prices, trade, recipes.
SECRET
He lets petty thieves go when they are hungry.

Édouard

NPC 02
ROLE
Farmer. Five generations on the same land, three children.
PERSONALITY
Humble, honest, observant. He says less than he sees.
WHAT HE KNOWS
The seasons, the villagers, the local rumors.
WHAT HE DOESN’T KNOW
The Baron’s politics, market prices, faraway cities.
SECRET
He can read, and hides it for fear of being mocked.

Margot

NPC 03
ROLE
Merchant for twelve years. A widow with a twelve-year-old son.
PERSONALITY
Charming, calculating, never malicious. Always three moves ahead.
WHAT SHE KNOWS
Trade routes, prices, rumors, the account books.
WHAT SHE DOESN’T KNOW
Liturgy, military tactics, the secrets of the forest.
SECRET
She sometimes buys goods that “fell off the cart”.
Interactive scene · Bertrand, guard of Vertpré

Choose what you do.

Reference replies from the character sheet
1 · ENGINE DECISIONDETERMINISTIC
SEEN
player: “Hi guard, I'm looking for an inn for the night.”
RULE
role: help travelers
INTENTION
aider
RELATION
trust 0.30 · starting value
2 · NPC LINELOCAL LLM

BERTRAND, looks you up and down The Boar Inn, at the end of the street. Tell Gérard that Bertrand sent you. And... stay out of trouble, alright?

In-game dialogue is written in French. Translated here.

The real numbers

Base 100 = previous situation
Delay before the NPC starts replying
0.6 sabout 5× faster

Half of all exchanges start in under 0.6 s, versus 2.9 s before.

The reply appears as it is written.

Median of 59 exchanges, measured Oct 19, 2026 on a mainstream laptop, development version of the game.

Measurement details

19 exchanges in 20 start in under 1.1 s when the reply is progressive (estimate from 50 exchanges).

About 1 exchange in 7 appears all at once, in 2.9 s (special cases).

The first exchange with an NPC takes about 1 s.

No stutter longer than 0.1 s across 88 measured exchanges.

Well-structured model replies
≈ 98%

readable by the game engine on the first try

Measured Oct 19, 2026: 86 of 88 replies in play, 98.5% across 3,000+ test replies.

Fine-tuning dataset
924examples
86/86 INTENTIONS
124/124 REACTIONS
Target hardware
6 GBVRAM · RTX 4050

Fine-tuned, quantized Llama 3.2 3B runs alongside the Unity game, on the same card.

Project status

  1. TODAY · REACHEDPlayable proof of concept
  2. EARLY 2027 · UPCOMINGPrivate demo
  3. FALL 2027 · UPCOMINGFundraising
04 / Team

Five people, one village that stands on its own.

05 / What we’re looking for

Pre-train our own LLM, built to run alongside a game.

Public market studyEarly 2027

We will publish it here, open access. To get it early or to talk about the raise, write to us.

Contact us
06 / Contact

Let’s talk about Vertpré.

Investors, publishers: the prototype runs. We can show it to you live, on the machine of your choice.

Or write directly
contact@halcya-studio.com

The form needs JavaScript. Write to us directly at contact@halcya-studio.com.

FORM* REQUIRED FIELDS
You are *

Your information is used only to reply to you. It is emailed to the Halcya Studio team and is not stored by the site.