Study Finds WikiTrace Scores 8.5 Points Above Prior Agent Memory Method

Researchers Yukyeong Lee, Yumin Choi, Jinheon Baek and Sung Ju Hwang developed WikiTrace, a memory system for AI agents that links stored lessons to evidence from the tasks that produced them. The system scored an average of 8.5 points above the strongest prior memory method across five agentic benchmarks.

AI agents use external memory to apply lessons from past tasks. But a lesson can be incorrect or applied in the wrong setting, causing an agent to fail a task it could have completed without memory. This failure mode is memory-induced degradation.

WikiTrace connects each lesson to a summary explaining why its original task succeeded or failed, as well as the task’s raw trajectory. Agents can use those records to check and refine lessons before acting.

The evaluation found that WikiTrace consistently outperformed existing memory approaches and a baseline using the same insights without evidence links. It also had the lowest regression rate, 4.5%, across the five benchmarks.

The paper, “Don’t Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents,” was published on arXiv Nov. 10. An open-source implementation is available on GitHub. TokenPost has also covered a separate memory agent’s results on Terminal-Bench 2.0.