According to God of Prompt on X, IBM introduced Trajectory-Informed Memory (TIM), a method that observes an agent’s full execution trace and extracts reusable guidance—what worked, what failed and how it recovered, and what succeeded but wasted steps—to inject into future prompts for similar tasks, with the base model unchanged and no retraining required. As reported by the post, TIM delivered a 14.3 percentage-point gain in scenario completion on unseen tasks and lifted complex task completion from 19.1% to 47.6% (a 149% relative increase), targeting 50+ step, multi-application workflows where agents commonly fail. According to the same source, the business impact is lower iteration costs, faster time-to-value in production agent deployments, and safer rollouts by encoding recovery strategies directly into prompts, creating a practical path to scalable, memory-augmented agents without model fine-tuning.