Memory

In short: Many AI agents look productive but are actually drifting — confidently executing the wrong moves on a wrong picture of the situation. The bottleneck for the next phase of agent systems is not larger context windows or stronger base models; it is whether the system can construct and maintain a stable belief state. This piece argues why belief state quality is the right optimization target, proposes five proxy metrics to measure it, and lays out where to put incremental engineering resources next. AI agents that look productive often turn out to be drifting — confidently executing the wrong moves on a wrong picture of the situation. Competition in agent systems is shifting from “whose model is stronger” toward “who can keep producing higher-quality belief state.” If you accept that framing, several seemingly unrelated problems suddenly line up: the same model behaves very differently inside different product shells; long-running agents fail not because they cannot answer but because their judgment of the situation is wrong; context windows keep growing, but system capability does not scale linearly with them; and scattered engineering pieces — skill, memory, retrieval, tool use, trace, summary — all start to matter at the same time. ...

Just a year ago, people compared reading lists and book recommendations. This year, nearly every conversation seemed to revolve around AI. There is no denying that AI is an extraordinarily powerful tool. But convenience has a cost. As reliance grows, thinking quietly recedes. As one widely circulated line puts it: “We are trading depth of thought for speed of AI.” A growing body of research suggests this trade-off is real. When MIT Researchers Sound the Alarm Some of the earliest and most serious warnings about AI dependency have come not from skeptics, but from researchers at the forefront of the technology itself. At the MIT Media Lab, research scientist Natalia Kosmina led a striking experiment examining what happens inside the brain when complex cognitive tasks—like writing—are outsourced to AI. Her team recruited 54 undergraduate students from institutions including Harvard, MIT, and Wellesley College. Participants were asked to write SAT-style argumentative essays under three different conditions: Brain-only group: no external tools Search group: access to Google AI group: access to ChatGPT Throughout the task, all participants wore EEG devices to monitor real-time neural activity. ...