Programmatic Memory for Long-Horizon LLM Agents
PRO-LONG enhances long-horizon LLM agents' performance by integrating programmatic memory.
PRO-LONG introduces a minimal memory addition for LLM agents in long-horizon tasks. It logs every observation, action, and outcome to a structured log file, allowing agents to retrieve and reason over this data programmatically. This approach yields an 18 percentage point improvement over coding agents without the log.