WikiTrace links agent lessons to trajectories to cut memory errors
Original title: Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
Agents can descend from insights to trajectory summaries and raw traces before acting when lessons conflict or lack detail.
The paper describes memory-induced degradation, where LLM agents fail tasks they would solve without memory because they follow wrong or misapplied lessons. WikiTrace stores each lesson as an insight linked to a success-or-failure summary and the raw trajectory. On a new task, the agent starts from insights and uses terminal tools to fill gaps or correct conflicts before acting. The authors say it beat existing memory methods and an insights-only baseline on five agentic benchmarks while reducing that degradation.
Unconfirmed: Benchmark gains and reduced degradation are author-reported; the five benchmarks are not named.
- Lessons distilled only from success or failure labels are often given without their source trajectories.
- WikiTrace keeps the insight, a why-summary and the raw trajectory together.
- Code is released at github.com/yukyeongleee/wikitrace.
From the original article
1 min read at huggingface.co
External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks. However, memory can also lead to memory-induced degradation, where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one.
By Yukyeong Lee, Yumin Choi, Jinheon Baek · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.