KARMA integrates long-term and short-term memory modules into embodied AI agents, significantly improving their ability to execute complex, interconnected household tasks.

• Overcomes in-context memory limitations in traditional LLMs
• Distinguishes between long-term knowledge and short-term task information
• Uses memory-augmented prompting to enhance planning capabilities
• Reduces inefficiencies and errors in sequential task execution

For Engineering teams, KARMA represents a crucial advancement in creating more capable autonomous agents that can maintain context across extended operations—mimicking human-like memory capabilities essential for real-world applications.

KARMA: Augmenting Embodied AI Agents with Long-and-short Term Memory Systems