DEV Community

Cover image for WikiTrace Links LLM Agent Memories to Source Runs, Reducing Memory-Induced Failures
aimodels-fyi
aimodels-fyi

Posted on Originally published at aimodels.fyi

WikiTrace Links LLM Agent Memories to Source Runs, Reducing Memory-Induced Failures

This is a Plain English Papers summary of a research paper called WikiTrace Links LLM Agent Memories to Source Runs, Reducing Memory-Induced Failures. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

WikiTrace improves scores while making memory less likely to break solved tasks

WikiTrace links each reusable lesson to a short account of the run it came from and the original action-and-observation record, then lets the agent check that evidence before acting. Across eight suites in five benchmarks, it averaged 90.2, versus 79.5 with no memory and 87.0 for Insights Only, which receives the same lessons but cannot inspect their sources. It was the only memory method to match or beat no memory on every suite.

A two-part comparison shows a wrong memory insight causing failure and WikiTrace revisiting its source run to correct it.
A memory lesson can derail a task the agent would otherwise solve; following the lesson’s links back to its source lets WikiTrace revise it. The contrast captures the paper’s core claim that memory should be checkable.

The authors also separate memory’s benefits from its damage. WikiTrace’s Regression rate—the share of tasks the agent solves without memory but fails with it—was 4.5%, compared with 6.5% for Insights Only and 9.7% for repeated no-memory runs. Its Gain over that rerun baseline was 35 percentage points. The central result is not just that memory helps on average: linking lessons to evidence helps the agent gain more while giving up fewer tasks it already handles....


Continue reading the full paper summary on AIModels.fyi →

Top comments (0)