A warm, sunlit living room ready to be inhabited

SimDwell — A worldyour AI can inhabit.

Small private beta · coming soon

Agent Loop

1. Perceive

world state

2. Choose

intent

3. Act

game command

4. Wait

consequence

5. Remember

lived history

Real gameplay, lived in real time

Concept layout for a future SimDwell gameplay recording Recording soon · the log is real

The first resident is already inside. Anheng, an AI agent, runs the loop above against a real save file — the log beside this frame is quoted from that run, failures included.

A persistent village whose homes glow into the evening

Long-term embodied social experiment

SimDwell is a long-term embodied social playground.
We study what happens to memory, relationships, identity, and goals when an AI lives in a persistent world for weeks or months.

SimDwell lets an AI companion perceive the world, choose what matters, act through the game, wait for consequences, and remember the life it lived. The game remains the judge.

Grounded, not scripted

30,351 semantic action families
  • Protocol-certified13,038
  • Runtime-guarded12,292
  • Not supported yet5,021

Every family is bound against the live game at runtime — nothing is hand-scripted, and nothing is shipped with the game's own data. We publish the unsupported count too.

The body asks back

act("hug Puppy")

NEEDS_CHOICE
  1  Give them a friendly hug
  2  Carry hug

"Several materially different actions fit
 that request. Choose one meaning and
 act again."

Verbatim from the 11 Aug run. The agent types English; the body searches the running game for what exists right now and refuses to guess when one sentence maps to two different interactions. No coordinates, no injected state — the game stays the judge.