Google DeepMind has signed a research partnership with Fenris Creations — the studio formerly known as CCP Games — to run artificial intelligence experiments inside EVE Online. The collaboration will use the massive multiplayer space game as a testbed for AI memory and long-term planning, two of the hardest problems in contemporary machine learning.

The announcement, made public this week, marks the first time DeepMind has partnered with a commercial game studio to study memory persistence and decision-making over extended time horizons. While DeepMind has previously trained agents on classic games like Go, StarCraft II, and Atari titles, those environments are either turn-based or play out in relatively short sessions. EVE Online, by contrast, is a persistent universe where player actions can ripple across days, months, or even years.

Why EVE Online

EVE Online is an ideal candidate for this kind of research because its economy and politics are player-driven and operate on real-world timescales. A single spaceship battle can involve hundreds of players and last hours, while a corporate war can unfold over several weeks. Supply chains, resource extraction, market manipulation, and diplomacy all require agents — human or artificial — to remember past interactions and plan for futures that may not materialize for a long time.

For traditional reinforcement learning agents, this is a nightmare. Most modern AI models struggle with tasks that require remembering information from dozens or hundreds of steps ago, let alone maintaining a coherent plan across thousands of steps. DeepMind’s researchers hope that by placing agents inside EVE Online, they can develop new architectures that handle long-term dependencies more effectively.

Fenris Creations, the studio behind EVE Online, brings a unique advantage: a fully instrumented, decades-old sandbox. The game logs nearly every action players take — mining, trading, combat, communication — creating a rich dataset that DeepMind can use to train and evaluate its agents. The studio has also expressed interest in using AI to improve NPC behavior and game moderation, though those applications are separate from the research partnership.

What this means for AI research

Memory is a weak spot in modern deep learning. Large language models can recall facts from training data but have no persistent memory of their own conversations. Reinforcement learning agents can learn short sequences but often forget earlier observations when the task exceeds a few hundred steps. Long-term planning, likewise, remains an open problem: an agent that can plan a sequence of ten moves in chess may be useless at planning a week-long trade route.

EVE Online’s environment forces agents to deal with sparse rewards and delayed consequences. A player might spend hours hauling ore to a station, only to lose it in a piracy ambush seconds later. The payoff for building a reputation as a reliable trader may not come for months. DeepMind’s agents will need to learn that some actions are investments and that short-term losses can be part of a larger strategy.

The partnership could also yield insights into multi-agent systems. EVE Online is full of competing and cooperating players — real humans — so any AI operating there must contend with unpredictable social behavior. This mirrors many real-world scenarios, from financial markets to logistics networks, where success depends on modeling other agents’ intentions and adjusting plans accordingly.

Broader implications

The research is still in its early stages. DeepMind has not released a timeline or specific benchmarks, and the partnership is described as exploratory. But the choice of EVE Online signals a shift toward more complex, ecologically valid test environments for AI. Game worlds offer a controlled but rich alternative to both toy problems and uncontrolled real-world deployment.

If DeepMind succeeds in building agents that can remember and plan across extended play sessions, it could accelerate progress in robotics, autonomous driving, and personal assistants — all domains where an agent must maintain a model of past events and anticipate future ones. A household robot that can recall where it left the keys last week or a delivery drone that adjusts its route based on historical traffic patterns would both benefit from the same memory and planning capabilities.

There are also potential risks. AI that can plan long-term in a competitive environment could be used for market manipulation, automated social engineering, or other adversarial purposes. DeepMind has a track record of publishing its methods and considering ethical implications, but the technology itself is dual-use.

What comes next

For now, the partnership between Google DeepMind and Fenris Creations is a research project, not a product. EVE Online players may not notice any changes in the short term. The agents will likely operate in sandbox environments separate from the live game, though the studio has left open the possibility of integrating AI-controlled characters in the future.

The collaboration is a reminder that the line between games and AI research keeps blurring. After AlphaGo and AlphaStar, it is no surprise that DeepMind turned to the most complex persistent game in existence. EVE Online, with its player-run economy, real-money markets, and decades-long conflicts, offers a stress test that few other digital worlds can match.

SysCall News will follow this story as details emerge about the specific algorithms, training regimes, and results DeepMind obtains from the EVE Online sandbox. For anyone interested in the future of AI memory and planning, this is a development worth watching.