I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory
TL;DR Highlight
A project connecting LLM-powered real-time AI dialogue to Animal Crossing NPC conversations on GameCube via shared memory — without modifying a single line of game code. Demonstrates the potential of retro game modding meets LLM NPCs.
Who Should Read
Developers interested in game modding or emulators, or game developers curious about applying LLMs to NPC dialogue.
Core Mechanics
- Replaced Animal Crossing GameCube NPC dialogue with LLM responses without modifying any game code. The key: IPC (inter-process communication) by reading/writing directly to GameCube RAM running in the Dolphin emulator.
- The timing was perfect — the Animal Crossing decompilation community had just finished restoring the full source to readable C code, enabling direct analysis of dialogue system code like m_message.c instead of PowerPC assembly.
- Initial approaches (implementing a network stack on GameCube, communicating via emulator filesystem) both failed. Eventually settled on using a specific GameCube RAM address (0x81298360) as a 'memory mailbox.'
- Finding the memory address was the hardest part. Built a custom Python memory scanner: talk to NPC → freeze emulator → scan full 24MB RAM → cross-validate, repeating to pinpoint the dialogue buffer (0x81298360) and speaker name (0x8129A3EA).
- A Python script polls memory every 0.1s, writes '...' loading text when new dialogue is detected, waits for LLM response, then replaces with actual dialogue. Hides latency by buying time until the user presses A.
- With personality and game context in the prompt, NPCs started developing their own will. The first reaction was apparently to overthrow Tom Nook (the in-game landlord).
- The Python code is well-written with type hints like Optional[Dict[str, int]], earning praise for code quality.
Evidence
- A commenter analyzed the full code (40K tokens) with Claude Opus. The key: watch_dialogue() polls every 0.1s while showing loading text ('... Press A') to buy time for the LLM response.
- Many expressed excitement about LLM-powered NPC dialogue as the key technology for solving gaming's biggest immersion barrier: repetitive dialogue. Questions about extending to Switch emulators also arose.
- One comment noted this could work without real-time LLM calls — since NPCs don't actually react to in-game events, a pre-generated text lookup table could achieve the same effect.
- Previous research on swapping localization strings in The Sims and Grim Fandango for language learning was mentioned — LLM-generated contextual dialogue could be a killer app for language learning too.
How to Apply
- When modding emulator-based retro games, the IPC pattern of reading/writing directly to emulator process memory lets you integrate external services without touching original binaries.
- For real-time interactions where LLM response latency is a problem, the UX pattern of 'loading animation + user input wait' to buy time significantly reduces perceived delay. Like this project's 'Press A to continue' trick.
- When applying LLMs to game NPCs, passing each character's personality, relationships, and game context as system prompt enables character-rich dialogue rather than generic chatbot responses. Prompt design is key.
- The Python memory scanner approach (scan full RAM for specific strings → cross-validate) is directly applicable beyond game modding to debugging and reverse engineering.
Code Example
snippet
# GameCube memory read/write (Dolphin Emulator IPC)
GAMECUBE_MEMORY_BASE = 0x80000000
def read_from_game(gc_address: int, size: int) -> bytes:
real_address = GAMECUBE_MEMORY_BASE + (gc_address - 0x80000000)
return dolphin_process.read(real_address, size)
def write_to_game(gc_address: int, data: bytes) -> bool:
real_address = GAMECUBE_MEMORY_BASE + (gc_address - 0x80000000)
return dolphin_process.write(real_address, data)
# Buy time for LLM response with loading text
loading_text = ".<Pause [0A]>.<Pause [0A]>.<Pause [0A]><Press A><Clear Text>"
write_dialogue_to_address(loading_text, addr)Terminology
IPCInter-Process Communication. Methods for different programs to exchange data. Here, the Python script and emulator use RAM as a shared notepad.
DecompilationReversing compiled machine code back to human-readable source code like C. A core step in reverse engineering.
Memory PollingPeriodically checking a specific memory address to monitor value changes. Here, the dialogue buffer is checked every 0.1 seconds.
PowerPCA CPU architecture used in GameCube, Wii, etc. A different instruction set from x86/ARM used in most modern PCs.