Meta AI has introduced a new memory module that pairs an action agent with a separate memory agent to help AI agents stay on track during long tasks. The system aims to address common issues such as forgetting constraints, repeating failed commands, and rediscovering errors. The memory agent reviews recent steps and updates a structured memory bank, deciding whether to add a reminder to the action agent's next call or remain silent. This approach is designed to enhance the agent's ability to maintain context and avoid repeated mistakes. Source: thedecoder
The proposed system works by pairing an unmodified 'action agent' with a separate 'memory agent.' At fixed intervals, the memory agent reviews a sliding window of the most recent steps and updates a structured memory bank. It then decides whether to add a brief reminder to the action agent's next call or stay silent. The memory bank has three sections: a private status field, knowledge memory, and procedural memory. The system ensures that the action agent can only update the memory bank through predefined tool calls, rather than freely overwriting its contents. Source: thedecoder
The researchers tested the system on Terminal-Bench 2.0 and tau2-Bench, achieving improvements in task success rates. With the older Claude Sonnet 4.5 as the action agent, the system solved 46 percent of Terminal-Bench tasks on its first attempt, compared to 38 percent for the baseline. On tau2-Bench, the task-weighted average rose from 55 to 62 percent. The gains varied by domain, with improvements of about 10 percentage points in airline and retail tasks, and 3 points in telecom. Source: thedecoder
