Agent frameworks & orchestration
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
An arXiv abstract introduces WikiTrace, which lets LLM agents check stored lessons against the trajectories they came from.
Why it matters: For agent builders, WikiTrace makes checking the provenance of a memory lesson part of using it, rather than treating the lesson as sufficient on its own. The abstract’s reported results suggest this approach can improve performance while reducing memory-induced degradation, though it provides no effect sizes.