Google DeepMind partners with EVE Online studio to test agents that learn over months
Persistent game worlds are becoming testbeds for the memory and planning abilities that AI agents still lack.
Google DeepMind has announced a research partnership with Fenris Creations, the studio formerly known as CCP Games and the developer of EVE Online, in a blog post published on 21 August. The collaboration spans three environments: EVE Online, EVE Vanguard and EVE Frontier. The post, written by Alexandre Moufarek and Adrian Bolton, frames the work as building on 15 years of research in games that began with agents learning to play Atari titles.
The partnership targets four capabilities the company describes as frontier problems for agents: continual learning in worlds that change, long-term memory that extends beyond a model's context window, long-horizon planning over weeks, months or years, and complex multi-agent dynamics involving cooperation, competition and emergent behaviour. The work will start in an offline instance of EVE Online used as a sandbox, move to EVE Frontier to study how humans and AI coexist, and eventually bring mature capabilities into the live EVE Online and EVE Vanguard games.
Several DeepMind systems are involved. SIMA 2, described as a generalist agent powered by Gemini, follows natural-language instructions, perceives gameplay the way a player would and acts through standard keyboard and mouse controls. Genie 3 generates interactive worlds, and an existing Aura Guidance feature in EVE Online already uses Gemini to surface player-generated knowledge for newcomers.
Fenris Creations chief executive Hilmar Pétursson said EVE Online was designed as a sandbox of lasting consequences shaped by its players, and that the collaboration pushes into territory where AI must learn, adapt and remember over timescales no other game demands. DeepMind also lists partnerships with studios including Coffee Stain Studios, Hello Games, Keen Software House, Strange Loop Games, Thunderful Games, Digixart and Tuxedo Labs.
Why it matters
Benchmarks that reset after each episode cannot measure whether an agent remembers, plans or adapts over long periods, and those are exactly the abilities that separate a demo from a dependable assistant. A persistent online world with thousands of human players offers consequences that carry forward for years. Choosing such an environment signals that the research frontier has shifted from winning games to surviving in them. The staged rollout also shows how labs are trying to test agents safely before exposing real users to them.
Offline EVE Online instance
A private copy of the game used as a safe sandbox for early experiments.
EVE Frontier
A separate environment for studying how humans and AI agents coexist.
Live EVE Online and EVE Vanguard
Mature capabilities eventually integrated into the games players use.
What you can learn from this
An instructable agent closes the loop from pixels to actions. SIMA-style agents take a screen image and a text instruction as input and produce keyboard and mouse events as output, the same interface a human uses. That design means the agent needs no special hooks into the game and can in principle transfer to any software with a screen, from a different game to a spreadsheet. It also makes the problem hard, because the agent must infer game state from what it sees rather than reading it from memory.
A context window is working memory, not long-term memory. A language model can only attend to the tokens currently in its context, so anything that happened earlier is gone unless it is written down somewhere and brought back. Long-term memory for agents therefore means external stores, retrieval systems and summaries that decide what is worth keeping. The engineering question is not how large the window is, but what to store, when to recall it and how to keep it accurate as the world changes.
Continual learning fights catastrophic forgetting. When a neural network is trained on new data, updating its weights tends to overwrite what it learned before, a failure known as catastrophic forgetting. Most deployed models avoid the problem by freezing weights after training, which is why they cannot learn from experience. A world that changes over months forces researchers to confront how an agent can improve without losing its earlier skills.
Long-horizon planning is about credit assignment. In reinforcement learning, an agent learns by connecting rewards to the actions that caused them. When the payoff arrives weeks after the decision, the chain of intermediate steps is enormous and the signal is weak. Techniques such as hierarchical goals, sub-tasks and learned value estimates exist to bridge that gap, and a game with economies and alliances spanning years is a demanding place to test them.
Multi-agent environments produce behaviour no one programmed. When many agents, human or artificial, pursue their own goals in a shared world, alliances, markets and conflicts emerge from their interactions. Studying these dynamics matters because real deployments will involve many agents acting alongside people, and an agent that behaves well in isolation may behave very differently when others compete with it. Starting in an offline copy of the game is a way to observe emergent behaviour before it can affect real players.
We teach this
How to use this in practice
Give a chat assistant a persistent memory file. Write a small script that wraps calls to any language model API and, after each conversation, asks the model to append three durable facts about the user or task to a local text file. Load that file into the system prompt at the start of the next session, and add a rule that contradictory facts replace older ones rather than accumulating. Done looks like the assistant correctly recalling a preference you stated yesterday without it being in today's transcript, and updating it when you change your mind.
Run a classic reinforcement learning environment. Install the Gymnasium library and train a simple agent on CartPole or a small grid world using a tutorial implementation, logging the reward per episode. Done looks like a plot showing reward rising over a few hundred episodes, which gives you a concrete feel for what learning from consequences means before you read about agents in complex worlds.
Test forgetting with a long conversation. Start a session with any chat model, state a specific fact early, then fill the conversation with unrelated content until you exceed the model's context length, and ask about the fact. Try it a second time with the fact restated halfway through to see how position affects recall. Done looks like a note recording where recall failed and a short design for a retrieval step that would have fixed it.
Draw a memory architecture for an agent. Sketch the boxes for perception, working context, an episodic store of past events, a semantic store of stable facts and a planner, then draw the arrows showing when information moves between them. Label each arrow with the trigger, such as end of session or a failed lookup, that causes the transfer. Done looks like a one-page diagram you could use to explain why context size alone does not make an agent remember.
Sources
- From Atari to EVE Online: Building on 15 Years of AI Research in Games — Google DeepMind
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report