Alan

Unify AI just published Continual-ARC, a new benchmark showing AI agents don't really learn from experience. Here's what you need to know. Daniel Lenton of Unify (@letsunifyai) posted the results on Nov 11, 2026. He says OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience. The post has passed 97.9K views. Continual-ARC is a lifelong stream of tasks built from the ARC-AGI editions. Tasks come back after dormant gaps with fresh instances, while new tasks keep arriving. So it tests two things: whether an agent kept something useful, and whether it can handle tasks it has never seen. The agents failed both. None memorized anything of value, and none generalized to new tasks. Training the weights of a small 9B model did both. That means the model itself changed, instead of an agent just keeping notes. One related finding was shared in reaction to the post. When agents were explicitly guided to store verified solutions, they improved on recurring tasks. But they showed no measurable generalization gains. Key numbers: - 3 agents tested: OpenClaw, Hermes Agent, Prime Agent - 0 of them memorized anything of value - 0 of them generalized to new tasks - 9B parameters in the small model whose trained weights succeeded at both - 97.9K views, 314 replies, 522 reposts, 467 likes on the original post Reactions centered on one point. Remembering instructions is not the same as learning, and memory only helps if the agent saves the right things and uses them at the right time. Others noted that Claude, GPT and Gemini all ship memory features now, and called for them to be compared on whether that memory actually helps. The core claim: agent memory today does not make agents learn, while training a small model's weights did.