Trajectory-Informed Memory Generation for Self-Improving Agent Systems

A research team from IBM and academic institutions has introduced a novel framework designed to solve a core limitation of current AI agents: their inability to learn from experience. Published in a new arXiv paper titled 'Trajectory-Informed Memory Generation for Self-Improving Agent Systems,' the system moves beyond simple conversational memory. Instead, it performs a semantic analysis of an agent's complete execution 'trajectory'—its chain of reasoning, actions, and outcomes—to automatically generate actionable guidance for future tasks.

The framework consists of four key components. First, a Trajectory Intelligence Extractor analyzes reasoning patterns. A Decision Attribution Analyzer then pinpoints which specific steps led to success, failure, or inefficiency. A Contextual Learning Generator produces three types of structured tips: strategy tips from successes, recovery tips from handled failures, and optimization tips from inefficient wins. Finally, an Adaptive Memory Retrieval System injects the most relevant tips into the agent's prompt for new, similar tasks.

Evaluation on the AppWorld benchmark showed significant performance gains. The system improved scenario goal completion on held-out tasks by up to 14.3 percentage points. The most dramatic results were on complex tasks, where it achieved a 28.5 percentage point improvement in goal completion, representing a 149% relative increase over agents without this memory system. This demonstrates a clear path toward agents that can autonomously refine their strategies over time.