Memory is where the idea of a “living game” gets difficult. An AI character can sound convincing for five minutes, but remembering what you did days or weeks ago is a different problem. Google DeepMind’s latest work with EVE Online is aimed directly at challenges like that.
We recently looked at Google Cloud’s argument for living games, where characters remember interactions and worlds react over time. DeepMind is approaching that broad idea from the research side. Its EVE work looks at decisions that unfold over long periods, persistent memory, continual learning, and interactions between multiple agents.
SIMA 2 exposes one of the biggest hurdles. DeepMind says the agent currently has relatively short interaction memory because it uses a limited context window to maintain low-latency responses. That makes EVE an interesting test case, but it also makes clear how far persistent AI still has to go.
EVE Online Never Stops Changing
EVE Online has been running since 2003, with thousands of people sharing the same persistent universe. Its economy, alliances, conflicts, trade networks, and political landscape have changed through more than two decades of human activity.
That is very different from giving an AI a short challenge and asking it to reach a goal. DeepMind wants agents to keep learning without forgetting what came before. They also need to retrieve information from far beyond the context windows used by current models. On top of that, an agent may need to make decisions whose consequences don’t become clear until much later.
EVE also adds other people and agents to the equation. Agents may have to deal with economic choices, negotiation, cooperation, competition, and unexpected behaviour from others.
A persistent world therefore creates a very different memory problem. Remembering an earlier conversation is one thing. Remembering enough of a changing world to make that conversation useful weeks later is much harder.
SIMA 2’s Short Memory Is Tied to Low-Latency Interaction
SIMA 2 already goes beyond the idea of an AI following a list of scripted commands.
The original SIMA learned more than 600 language-following skills across different 3D environments. It sees what is happening on screen and controls the game through a virtual keyboard and mouse. It doesn’t require access to the game’s source code or a special API. SIMA 2 adds Gemini-based reasoning, conversation, and the ability to improve through additional experience.
Its limitations are just as interesting. DeepMind says SIMA 2 still struggles with very long, complicated tasks that require many steps. It also has relatively short memory because keeping the context window limited helps the system maintain low-latency interaction. Reliable keyboard and mouse actions, along with visual understanding, remain open problems as well.
That memory-versus-latency tradeoff gets right to the heart of a living game. An AI companion might need to remember how you treated it earlier, what happened somewhere else in the world, and what has changed since then. At the same time, you still expect it to respond quickly when you speak to it or something happens nearby.
A longer memory isn’t very useful if every interaction starts feeling delayed. A fast response isn’t enough either if the character continually forgets the history that was supposed to make it feel persistent.
Long-Horizon Planning Can Span Weeks, Months, or Years
DeepMind isn’t only interested in remembering old information. It also wants agents that can reason over weeks, months, or even years.
That expands the problem considerably. A short AI interaction can be evaluated almost immediately. Did the agent understand the instruction? Did it move to the right place, and did it complete the task?
Long-horizon behaviour is different. An action taken now might make sense only because of something that happened earlier, or because of a goal that won’t be completed until much later. The world may also change while the agent is working toward that goal.
Continual learning complicates the problem further. DeepMind wants agents to gain new skills without losing knowledge they already acquired. In a world like EVE, that means adapting as the environment itself keeps changing.
This is where persistent AI starts looking less like one smart NPC and more like an ongoing systems problem.
Our earlier living-games discussion looked at persistent state, AI processing, and the infrastructure needed to keep those experiences responsive. DeepMind’s research doesn’t establish how a commercial implementation would eventually be hosted. It does show why keeping useful context over long periods is a much harder problem than generating one convincing response.
DeepMind’s First EVE Experiments Are Separate From the Live Game
DeepMind isn’t dropping experimental AI agents into the existing EVE Online universe. Fenris Creations says the partnership begins with an offline version of EVE Online running on a local server. That controlled environment lets the teams test and evaluate models without affecting the live community.
DeepMind says the longer-term program will then move through EVE Frontier. Its programmable Smart Assemblies and open architecture create another environment where agents may need to adapt as the rules around them change. EVE Vanguard gives the research team a faster first-person environment for tactical decision-making.
Only after those capabilities mature would DeepMind consider bringing them into live EVE Online or EVE Vanguard. That distinction is important because this remains research, not a confirmed roadmap for AI agents joining the existing game.
EVE does already have a much narrower AI experiment in Aura Guidance. Introduced as a prototype in February 2026, it uses a bank of real Rookie Help questions and answers to assist newcomers. Responses can account for context such as location, ship, and recent activity, while uncertain questions are redirected to human Rookie Help. EVE says the heavier processing happens off-client in the cloud.
Aura isn’t SIMA, and it isn’t evidence that DeepMind’s experimental agents are already running inside EVE. It is useful as a contrast. Aura demonstrates that AI can support a persistent online game without attempting to become an autonomous part of that world.
Living Games Still Have a Hard Technical Problem to Solve
The more interesting part of DeepMind’s EVE work is that it doesn’t make persistent AI sound easy.
SIMA 2 can reason, converse, transfer knowledge between different environments, and improve through experience. DeepMind is also clear about what it still can’t do reliably. Long tasks remain difficult, memory is short, and low-latency interaction places limits on how much context the agent can keep active.
Those limitations put some useful boundaries around the living-games conversation. An NPC producing a convincing response is one challenge. Creating one that remembers enough of your shared history to make a later response meaningful is another. Doing that inside a world that keeps changing while still responding quickly raises the difficulty again.
EVE Online may be one of the best environments available for testing those limits because persistence is already part of its identity. The next major step for living games may not be a character that says something clever. It may be an agent that remembers enough of what happened before to make what it says next actually mean something.
As always, remember to follow us on our social media platforms (e.g., Threads, X (Twitter), Bluesky, YouTube, and Facebook) to stay up-to-date with the latest news. This website contains affiliate links. We may receive a commission when you click on these links and make a purchase, at no extra cost to you. We are an independent site, and the opinions expressed here are our own.
