Meta's new research: Your AI Agent Doesn't Forget. It Stops Listening to Itself.
Why the smartest models still repeat their own mistakes, and how a second agent that whispers "remember this" at the right moment beats bigger context windows
Give a modern AI agent a long, messy task and watch what happens. It reads the rules carefully at the start. It discovers, three minutes later, that a certain command fails. It diagnoses a bug with surprising skill. And then, twenty steps on while fixing something unrelated, it violates the rule it just read, reruns the command it just watched fail, and treats the bug it just diagnosed as brand new. Nothing was deleted. The information is still sitting there in the transcript, sometimes still inside the model’s context window. The agent simply stopped letting it shape the next move.
A team at Meta AI has a name for this quiet catastrophe, and a fix that is refreshingly modest in its ambition and surprisingly large in its payoff. The paper, “Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents,” reframes what memory is even for. The result is a plug-and-play module that lifts a frontier agent’s success rate by as much as 8.3 percentage points without touching the agent itself. That gap is the difference between a demo and a product.
