Most people who have run an agent on a long task, even on a laptop, have seen the same small failure. It works carefully for several steps, then repeats a command that already failed, or drops a constraint it was given at the start.
As the task runs, the useful detail gets buried under everything that came after. The agent still has it somewhere in its context, and behaves as though it does not.
In response to strong demand, we are running the NVIDIA AI Developer Bootcamp, an NVIDIA-certified workshop, in partnership with De La Salle University, Manila.
It runs three days and stays hands-on throughout:
1. Deep learning fundamentals. The core techniques and tools, worked through rather than lectured.
2. Data types and model architectures. Practice with the kinds you will actually meet at work, not textbook cases.
3. Transfer learning. Building a model by starting from one that has already learned the general patterns and adapting it to your own data, which is how most working models get built today.
Participants who complete all three days receive an NVIDIA certification.
When: 9 to 11 October 2026, Friday to Sunday, 9 AM to 6 PM
Where: De La Salle University, Manila
The root cause, as the researchers frame it, is attention. Giving the agent more memory does not fix it, because the relevant fact is present but no longer prominent at the moment it matters.
The common alternative. Feed everything back in at every step. The reasoning against it is that this is expensive and buries the important detail in an even longer pile.
The fix that worked. The Batch reported in September on Meta research describing a Proactive Memory Agent: a second agent whose only job is to watch the first. It keeps notes on facts, environment details, failed commands and unfinished tasks, then surfaces a reminder at the moment it becomes relevant. In the article's words, the aim is to highlight relevant memories "without overwhelming the agent's attention."
The measured effect, with Claude Sonnet 4.5:
• Terminal-Bench 2.0, 85 command-line problems: 37.6% to 45.9%.
• Tau-squared bench (written τ2-Bench), 278 customer-service problems: 55.0% to 61.8%.
An honest caveat. These are benchmark conditions, and the size of the gain varies by model and task.
Why it is worth knowing about. It needs no retraining and sits alongside an existing agent. Whether a team can try it directly depends on [confirm whether Meta released code].