AI That Remembers Its Own Mistakes Makes Fewer of Them

A team of researchers in Portugal has published a study on making AI "agents" — AI that can take actions for you, like clicking, booking, or filling in forms — less likely to fall apart halfway through a job. They took an existing agent called SwiftSage, which works like a two-speed brain: a fast part that suggests the next move, and a slower part that plans ahead. Then they bolted on two extra pieces: a memory module and a self-reflection module.

The memory module is like a notebook of moments that mattered, so the AI can look back when a similar situation comes up again. The self-reflection module is a mid-task gut check: before and after each action, the AI asks whether it's still on track, and corrects course if it isn't. The pair were tested in ScienceWorld, a text-based game where the AI must complete science tasks. The best setup finished 43% of tasks and scored 64.62 out of 100 — the strongest of four versions tested.

The surprise finding: the self-check did most of the heavy lifting on its own. Memory only started paying off once the AI stopped fumbling the basics. In plain terms, the bottleneck isn't that AI is dumb — it's that it loses the plot and can't recover when a step fails.

Why you should care: the gap between a chatbot that writes a nice paragraph and an assistant that reliably books your flight is exactly this — keeping track of where it is and recovering from errors. The catch: this is a lab experiment on a text game, not a real product, and it still failed more than half the tasks. Treat it as a signpost, not a launch.