Google DeepMind deepens EVE Online research partnership to push AI memory and long-horizon planning

Google DeepMind has detailed an expanded AI research partnership with Fenris Creations, the studio behind EVE Online, using the two-decade-old persistent space MMO as a testbed for frontier capabilities like long-term memory and multi-agent cooperation, alongside a wider roster of game-studio collaborations.

What's new

In an August 21, 2026 blog post, DeepMind wrote: "Our latest research partnership with Fenris Creations, the independent studio behind the EVE Universe, represents a new chapter in our history of AI research in games." The company describes EVE Online, launched in 2003, as "a massively multiplayer space simulation where thousands of players share a single universe that has evolved continuously for more than 20 years," with a player-driven economy and trade networks spanning thousands of star systems.

DeepMind says that scale and persistence make EVE a useful proving ground for four capabilities it considers essential for frontier AI: continual learning without forgetting prior skills in a constantly changing world; memory that accumulates and retrieves knowledge across timescales beyond current model context windows; long-horizon planning that reasons over weeks, months, or years; and complex multi-agent dynamics spanning cooperation, competition, negotiation, and emergent social behavior.

The research is built on SIMA 2, DeepMind's "Scalable Instructable Multiworld Agent," which the company says is powered by Gemini and "acts as an interactive companion capable of real-time reasoning and conversation." SIMA 2 perceives games the way a player would, following natural-language instructions through ordinary keyboard-and-mouse controls rather than an API or access to source code, and DeepMind says it has already achieved human-like play in No Man's Sky, Valheim, Hydroneer, and other titles.

The Fenris Creations partnership spans three distinct EVE environments: EVE Online's large-scale persistent universe, EVE Vanguard's first-person tactical gameplay, and EVE Frontier's programmable "Smart Assemblies" with open, extensible rules. DeepMind says the research program begins with an offline instance of EVE Online, "a safe sandbox, separate from live players," before progressing to EVE Frontier and, eventually, live environments. One concrete result has already shipped: the Aura Guidance system, which DeepMind says uses Gemini to surface player-generated knowledge from real Rookie Help questions and answers to new pilots.

Fenris Creations CEO Hilmar Pétursson is quoted in the post: "Together with Google DeepMind, we're pushing into uncharted territory where AI must learn, adapt and remember on timescales that no other game environment demands." DeepMind also names a broader set of game-studio partners beyond Fenris Creations, including Hello Games, Coffee Stain Studios, Foulball Hangover, Keen Software House, RubberbandGames, Strange Loop Games, Thunderful Games, Digixart, and Tuxedo Labs.

Context

DeepMind frames the EVE work as the latest chapter in a longer research lineage that runs from the Deep Q-Network's 2015 breakthrough playing Atari games from raw pixels, through AlphaGo's 2016 defeat of Lee Sedol, AlphaZero, MuZero, and AlphaStar's 2019 Grandmaster-level StarCraft II play. The company has previously said this game-driven research fed directly into AlphaFold's protein-structure predictions, work recognized with the 2024 Nobel Prize in Chemistry. SIMA, DeepMind's generalist gaming agent line, represents a shift from mastering specific games with a scoring objective toward general-purpose understanding of arbitrary game worlds.

Why it matters

Persistent, player-driven virtual economies like EVE Online offer something rare for AI research: environments that keep running and evolving for decades, with real scarcity, trade, and social dynamics that don't reset. That makes them a plausible stand-in for testing memory and planning horizons far longer than what fits in a model's context window today. If SIMA-style agents can genuinely learn, remember, and adapt inside a world like EVE without needing access to the game's underlying code, the same approach could generalize to other long-running, code-inaccessible software environments, an important capability for agents meant to operate autonomously over extended real-world tasks rather than single sessions.