# Will Wright’s Decade-Long Memory Game Gets a Cheaper AI Engine, Not a Bigger Budget

> Source: <https://rews.cc/a/will-wright-s-decade-long-memory-game-gets-a-cheaper-ai-engi-e6ee45>
> Published: 2026-11-18 18:10:33+00:00

**Key takeaway:** Splitting character behavior into a no-model reflex layer, an instinct layer and a planner that reserves model calls for reasoning-heavy decisions is how Artificial Agency's Behavior Engine keeps generative AI affordable at game-frame latency.

MyMems, the game Will Wright has been trying to ship since 2018, just changed its AI supplier, not its ambition. Gallium Studios, the ten-to-twelve-person studio Wright co-founded with Lauren Elliott, announced on November 18 that it will build the game’s characters and locations on the Behavior Engine made by Artificial Agency, a 20-person, Edmonton-based middleware firm, as [Ars Technica reported](https://arstechnica.com/gaming/2026/11/will-wrights-next-simulation-will-use-ai-to-mine-content-from-your-memories/). The pitch is the same one every middleware vendor makes to every studio that can’t afford to train its own models: let us run the inference, you keep the design.

The game asks players to type or dictate autobiographical memories, which a system called MemDNA turns into simulated places, characters and events. According to Artificial Agency’s own marketing for MyMems, MemDNA pairs a “Personal Behavior Model,” which grows as a player contributes more memories, with a “Knowledge Graph” mapping the people, places, events and relationships in those memories. “Meta knows what you click,” the company’s pitch reads. “Netflix knows what you watch. Nobody knows what you remember. MemDNA learns from your memories locally and only with your permission.” That “locally” matters: the model that builds the knowledge graph runs on the player’s device, not a Gallium or Artificial Agency server, which is the only thing standing between “game about your life” and “company with a database of your life.”

**How a memory becomes a game world** Gallium Studios' description of the MyMems pipeline

| From | To | How | 
|---|---|---|
| Player types/speaks a memory | MemDNA (on-device) *(Artificial Agency)* |  | 
| MemDNA (on-device) *(Artificial Agency)* | Personal Behavior Model |  | 
| MemDNA (on-device) *(Artificial Agency)* | Knowledge Graph *(people, places, events)* |  | 
| Personal Behavior Model | Behavior Engine *(reflex + instinct + planner)* |  | 
| Knowledge Graph *(people, places, events)* | Behavior Engine *(reflex + instinct + planner)* |  | 
| Behavior Engine *(reflex + instinct + planner)* | Simulated locale & characters |  | 
| Simulated locale & characters | Shared world *(family, friends, co-workers)* | explicit player action | 

Based on [Gallium Studios / Artificial Agency](https://www.unite.ai/gallium-studios-taps-artificial-agency-for-will-wrights-mymems/)

The Behavior Engine itself is the more interesting piece of engineering, and it has nothing to do with memories specifically. Per Artificial Agency’s technical documentation, [summarized by Unite.AI](https://www.unite.ai/gallium-studios-taps-artificial-agency-for-will-wrights-mymems/), the engine splits a character’s response into three layers that feed a single action list: a reflex layer that needs no model call at all (catching a thrown object, the documentation’s own example), an instinct layer, and a planner for decisions that need actual reasoning. That’s the standard trick for making generative AI affordable inside a loop that has to run at game-frame latency: push as much behavior as possible down to the cheapest layer, and reserve the expensive model call for the cases that need it. Wright put the economics plainly: the partnership gives Gallium “off-the-shelf” AI “without having to become an AI research company,” and avoids the “really expensive” frontier models he considers overkill for a ten-person studio’s game.

That line only makes sense against what Gallium has already burned through. The studio raised $6 million from Griffin Gaming Partners in 2022, on top of Wright’s own money, which the game-industry site VentureBeat and a Vulture profile this year both put at well over a million dollars of his personal funds plus additional millions from other investors. Gallium ran out in October 2024 and let its staff go; Wright has kept a skeleton crew working unpaid since, by his own account in that Vulture profile, because he’d “rather have a glorious failure than a mild success” than quietly shelve the project. Before Artificial Agency, Wright had been experimenting with a different AI backend: he discussed using Numenta’s NuPIC platform for the project at the GamesBeat Summit in May 2024, a detail that didn’t survive the pivot. Swapping AI vendors twice in two years is not what a studio with money to spare does.

The game itself predates the current AI boom by six years. Wright and Elliott announced it as Proxi: Yesterday’s You Tomorrow at the Game Developers Conference in 2018, after Wright says the concept had been nagging at him since 2016. The pitch was always the same: your memories become an avatar’s “brain,” which Wright told Vulture makes the whole project “kind of like Frankenstein.” Eight years, one funding round, one staff-wide layoff and at least two AI backends later, it has a new name, a new vendor, and still no release date more specific than 2027. VentureBeat quoted Wright on the practical problem that outlasted every technical pivot: investors keep passing because “it’s very hard to describe an idea as a designer,” and a ten-to-twelve-person team is, in his telling, a tough sell against publishers who reportedly want to see 20 million copies’ worth of upside before they’ll back a game.

Set against The Sims, Wright’s actual track record, the gap is stark. That franchise has sold more than 200 million copies and generated roughly $5 billion in revenue since 2000, according to VentureBeat’s reporting on EA’s own figures — a number no life-sim pitch deck can avoid mentioning and no AI memory game has come close to touching. MyMems is explicitly trading on that heritage; Gallium’s own marketing for the game invokes the 200-million-copy figure directly. Whether a knowledge graph built from a player’s own biography produces anything as repeatable as a procedurally generated suburban household is the open design question Wright has been sitting with since 2016, and the Artificial Agency deal answers how MyMems will run cheaply, not whether anyone will want to play it.

## What’s still unverified

Three claims in this story come from marketing copy, not measurement: that MemDNA’s on-device processing is secure against the kind of data leakage Meta and Netflix are implicitly being compared to; that the Behavior Engine’s reflex/instinct/planner split actually holds latency and inference cost down at the scale MyMems will need once an “early circle” of testers starts feeding it memories; and that 2027 is a real date rather than the same kind of placeholder Proxi has carried since 2018. None of the three has a public benchmark, a filing, or an independent test behind it yet. Artificial Agency’s $17 million CAD (about $12.5 million USD) seed round, led by Radical Ventures and Toyota Ventures and [reported by BetaKit](https://betakit.com/former-deepmind-researchers-secure-17-million-cad-to-bring-generative-ai-to-video-games-with-artificial-agency/), buys it runway to prove the engine on other studios’ games; MyMems is the first public case study, and the first one to run on a developer that has already gone through one full funding collapse.
