DeepMind Picks EVE Online as the Testbed for AI Memory That Lasts Years, Not Turns

Google DeepMind has partnered with Fenris Creations, the studio behind EVE Online, to use the EVE Universe — EVE Online, EVE Vanguard, and EVE Frontier — as a testbed for AI capabilities current models don’t have: continual learning without forgetting, long-term memory beyond today’s context limits, long-horizon planning across weeks or months rather than a single session, and multi-agent dynamics like negotiation and cooperation. The first player-facing result, Aura Guidance, uses Gemini to help new players learn the game; DeepMind plans to start with offline instances before considering integration with live servers. Fenris CEO Hilmar Pétursson frames EVE Online’s appeal for research as its design premise: a persistent sandbox “shaped by its players,” which demands an AI that adapts on timescales far longer than a benchmark task. The move continues 15 years of DeepMind using games as capability testbeds, from the original Deep Q-Network on Atari through AlphaGo.

Memory specifically — the capability DeepMind is targeting with EVE — is turning out to be less binary than “more context window is better.” A recent write-up describes IBM Research’s ALTK-Evolve system, which lets agents learn from past task trajectories by distilling guidelines at inference time rather than retraining weights. Tested across eight models from 30 billion to 745 billion parameters, the results split by model capacity: strong models gained the most from the full guideline set (DeepSeek-V3.2 improved 9.5 percentage points, Claude Opus 4.1 points), while weaker models did better with a selective, task-relevant subset — gpt-oss-120b gained 16.1 percentage points that way, using only 5% more tokens. Already-saturated models showed no improvement at all. IBM’s researchers summarized it as “agentic memory is not a feature you switch on. It’s a dose you calibrate to the model.”

Taken together, the two pieces point at the same open problem from different ends: DeepMind is building an environment complex enough to force long-horizon memory to matter, while IBM’s benchmark shows that once you have more memory, how much a given model can actually use it varies by an order of magnitude. Memory capacity and memory calibration are turning out to be separate engineering problems, and neither is solved yet.