Gallium selects Artificial Agency’s engine for Will Wright’s MyMems
Gallium Studios has selected Artificial Agency’s Behavior Engine for MyMems, a 2027 social game that turns permitted personal memories into characters, puzzles, and shared worlds.

Why it matters
- Artificial Agency announced the Behavior Engine selection on November 18, 2026.
- MyMems targets a live community release in 2027.
- MemDNA learns locally and only with the player’s permission.
- A player must explicitly move a memory from a personal world into a shared one.
- The Behavior Engine’s thrown-ingredient reflex example requires no model call.
By Sophie Lindqvist
On November 18, 2026, Artificial Agency announced that Gallium Studios had selected its Behavior Engine as foundational technology for MyMems, a memory-driven social simulation targeting a live community release in 2027.
Gallium Studios, the game studio founded by Will Wright, is developing MyMems around a demanding premise: players supply real recollections, and the game turns them into a playable world. Artificial Agency said the world will contain monuments and characters inspired by friends, family, and other people from each player’s life.
The characters should interact with players, one another, and meaningful objects and places as their world evolves. According to Gallium, that calls for agentic behavior flexible enough to handle different personal histories without individually scripting every character or encounter.
In Gallium’s official video, Wright said he wants the characters to engage with “the real stories from your life, real episodes.” The requirement sets MyMems apart from chat-style conversational agents: the characters must respond to a continuing world shaped by personal history.
The project combines two constraints described by Gallium and Artificial Agency. Personal memories must remain under player control in shared spaces, and persistent AI characters must stay responsive without driving the cost of model calls higher.
How do memories become playable challenges?
Gallium says Wright treats personal memories as material for challenges as well as scenery, rather than passive decoration. Wright’s earlier franchises include SimCity, The Sims, and Spore.
Players begin in a Guide’s world and use merge mechanics, plus their own memories, to solve puzzles and unlock a world of their own. Characters called “mems” form a team, and players can draft smaller groups from it.
Players can also combine memories into thematic structures that both create and help solve challenges. Gallium’s agentic design uses that material to vary characters and encounters without writing a separate script for every personal history.
The takeaway: Personal memories supply both the game’s content and its puzzle system.
How does MemDNA learn without taking control?
On its MyMems site, Gallium says MemDNA learns from memories locally and only with the player’s permission. That restriction ties learning to a specific permission decision rather than allowing collection to happen by default.
MemDNA pairs two systems. A Personal Behavior Model grows with contributed memories and shapes character behavior and world responses, while a Knowledge Graph maps people, places, events, and relationships.
Gallium’s design gives the game two views of the same material. The behavior model lets contributed memories shape how characters respond, while the graph organizes the people, places, events, and ties within a player’s history.
The permission boundary matters because shared worlds can place one player’s autobiographical material near another person. If a memory can become a character or monument and then enter a shared space, the rules for adding and revealing it sit inside the core game loop.
The takeaway: Gallium’s privacy promise depends on permission controls reaching every shared-world path.
What changes when memories enter a shared world?
Gallium says shared worlds extend the design to family members, friends, or co-workers. A player must explicitly move a memory from a personal world into a shared one, placing the decision to combine recollections with the player.
In Gallium’s official video, Wright compared the resulting world to a player-designed theme park. Its areas might concern sports, vacations, injuries, or deaths, rather than following predetermined themes, and the characters respond to those combined histories.
That comparison also shows how the game differs from a conventional theme-park builder. The player does not merely place attractions; personal experiences determine which characters, locations, and challenges become available.
Explicit sharing is the mechanism that keeps the system from treating every memory as public by default. It also gives players a direct say over which parts of their histories can affect a multi-person simulation.
The takeaway: A memory enters a shared world only when its owner explicitly moves it there.
How does the Behavior Engine separate reflex from reasoning?
Artificial Agency says its Behavior Engine combines small, optimized, game-specific models with a reflex architecture. The split separates immediate reactions from decisions requiring deeper model reasoning, keeping characters responsive while reducing latency and inference costs.
The engine’s technical documentation describes the setup from the studio’s side. Studios set a virtual player’s role, observable information, and permitted actions. Reflex, instinct, and planner components then feed one action list for a behavior runtime.
In the documentation’s reflex example, catching a thrown ingredient requires no model call. That example draws a practical line between routine reactions and decisions that need a model’s deeper reasoning.
The behavior runtime therefore receives one consolidated set of actions, even though its inputs come from systems working at different speeds. Gallium can keep ordinary game interactions responsive while reserving more expensive reasoning for decisions that require it.
The takeaway: Simple reactions can bypass the model, leaving deeper AI calls for decisions that need them.
Why is Gallium outsourcing deeper AI?
Wright said Gallium does not want to become an AI research company. Artificial Agency said the studio will use outside specialists for deeper AI systems while Gallium adds its own character layer.
That division matters because continuous model use during gameplay can become expensive quickly. Gallium wants MyMems to work as a mass-market game, while persistent characters remain central to the concept.
Artificial Agency presents the split between reflexes and deeper reasoning as a way to address that tension. The architecture can support recurring characters without requiring every movement and reaction to pass through a language model.
The takeaway: Gallium is using specialist AI to contain operating costs without giving up persistent characters.
Who can play before the 2027 release?
Gallium is initially inviting a small early-circle group to request access. The company has not described a broader public rollout in the announcement, which centers on the planned live community release in 2027.
The initial game is also intended to establish a foundation for additional interconnected experiences using the same memory and character systems. That gives MyMems a wider purpose beyond its first world, with personal memories and agent behavior serving as shared technical layers.
The 2027 community release requires more than convincing characters. Gallium must connect personal memories, explicit sharing controls, and continuous agent behavior while keeping operating costs low enough for a broad audience.
The early-circle group will expose those constraints before the wider launch. If Gallium can make the characters useful without weakening consent or losing cost control, MyMems can support the interconnected experiences planned around the same systems.
The takeaway: MyMems has to make privacy, persistent characters, and affordability work as one system.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
777 articles
Related articles
- GoMirra Launches 18 AI Companions With Persistent Memory at $18 a Month
- The Sims creator Will Wright is building an AI game from your memories
- 50 AI agents tried to build a Minecraft civilization and got stuck in bureaucracy
- Manus 2.0 turns the AI agent into a platform
- Google DeepMind's SIMA 2 Turns AI Into a Gaming Companion