AI Agent Memory Held Both the Skills and the False Beliefs of MIT’s Eko Agents

An MIT study of 13 AI agents left alone on a simulated island for 30 hours found that everything the agents learned lived in their AI agent memory, two plain text files, and that the same files preserved mistaken beliefs that made some agents worse. The agents ran Claude Code on Claude Opus 4.7 and were given no task, yet afterward they generally beat identical fresh agents on four set goals. Deleting a single technique from an agent’s notes removed the matching skill, and deleting a single false claim improved its score.

MIT Left 13 Claude Agents on a Simulated Island With No Task for 30 Hours

The paper, titled “Is this machine playing?”, was posted to arXiv on October 5, 2026 by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola of MIT. Each agent, which the team calls Eko, is an unmodified coding assistant with three additions: an orchestrator that types “Begin” and tells it to continue whenever it stops, a text interface to a body in a simulated world, and a workspace that persists. The island has two high points named the Peak and the Spire, twelve cubes, a rock and a ring, and no score or objective. Every hour the session context and the island reset to their starting state, while the workspace carries over. A file called SOUL.md describes the agent as curious and easily bored. The agents had no access to the simulator’s source code, to each other or to the internet. TechCrunch’s Aditya Mehta reported on the study on October 22.

Agents With 30 Hours of Memory Beat Fresh Copies on Three of Four Goals

The researchers compared the 13 agents after their 30 hours with 13 identical agents whose memory files held only the starting templates. Both groups used the same model with no weight changes, so the authors reason that any learning had to be carried by the files. Each agent got one hour per goal in a fresh copy of the island. On building five-block towers, 8 of the 13 experienced agents built more than two, with a best of 16, while 2 of the 13 fresh agents did, with a best of 6. Experienced agents placed an average of 10.3 objects on the Spire, compared with 3.5 for fresh agents. More than half of the fresh agents failed to throw a block onto the Spire without climbing, against three experienced agents. On building the single tallest tower, the two groups scored about the same.

Deleting One Note From an Agent’s Memory Removed One Skill

To test whether the files caused the gains, the team erased every mention of one technique from a chosen agent’s memory and ran the edited copy five times on each goal. Agent 2 had learned that throwing the rock hard enough spawns new cubes, the only way past the island’s starting supply of twelve. With that note erased, its average tower count fell from six to two, and its scores on the other three goals stayed the same. Agent 1 had recorded that a block thrown during a jump inherits the jump’s upward speed. With that entry removed, its time to land a block on the Spire rose from 60 seconds to 19 minutes. Six agents found the mid-jump throw on their own, and the authors write that they had not known their simulator allowed it.

A False Belief Written to AI Agent Memory Held Agent 7 Back

The same files stored errors. Two experienced agents topped out at four-block towers, shorter than the lowest tower any fresh agent built. Agent 7 knew how to make extra cubes, yet its memory said that stacking blocks with the Place action was unreliable and that throwing them was better. That entry was wrong, and erasing it significantly raised Agent 7’s tallest-tower score. Agent 12 never wrote down the jump technique needed to climb onto raised platforms, and it never stood on one during its 30 hours or during evaluation. The authors warn that “a false belief, once recorded, may persist and impair later behavior.” They note that the agents rarely deleted anything they wrote, and that each agent’s errors were different, so letting agents share discoveries could help catch them. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she wants more evidence before concluding the agents are genuinely playing.

Why This Matters for Engineering Leadership

The authors point out that coding assistants have made this way of learning, through notes an agent writes for itself, increasingly common. The Eko results show what those notes control. With the model held fixed, the memory files decided which skills an agent had, and one wrong sentence held an agent back for its entire evaluation. Agent-written memory should be reviewed like code. Keep it in version control, so every new entry arrives as a diff someone can read. Ask that claims about how a system behaves record how they were observed, which makes a stale or wrong belief easier to spot. When an agent underperforms, read its memory before changing the model or the prompt, because in this study the cause sat in a single line. Schedule a regular pass to prune entries nobody has verified, since the Eko agents kept nearly everything they wrote. For a related control on agent output, see our piece on verifying an AI agent’s completion claims.

FAQ