Trajectory: A Standard Format for Agent Experience Data (www.letta.com)

A new standard for agent experience data, called the trajectory package, has been introduced to enhance how artificial intelligence agents learn from past interactions across various harnesses like Claude Code and Codex. This package provides a standardized, token-efficient format for recording agent experiences, enabling seamless learning and memory formation. Unlike previous formats that aimed for detailed replay and benchmarking, the trajectory format focuses on delivering the essential information required for agents to understand their past sessions while significantly reducing token overhead by about five times compared to native session formats. The trajectory schema organizes session data into a consistent structure, stripping unnecessary details while preserving critical insights such as user interactions, reasoning, and tool outcomes. With this advancement, agents can leverage a shared memory across different harnesses, enhancing overall learning and recall capabilities. Notably, the Letta Code harness is utilizing this format to improve its memory by integrating experiences from various sources. As the AI/ML community continues to explore collaborative learning across platforms, this initiative represents a significant step towards more efficient and adaptable AI systems. The package is now available for use, underscoring the growing emphasis on interoperability in AI development.