Linking memory back to raw data helps AI agents avoid repeating bad lessons

The authors show that storing a path from each learned insight to its original trajectory lets agents catch and correct wrong instructions, reducing failures from 6.5% to 4.5%.

Top University

Yukyeong Lee · Yumin Choi · Jinheon Baek · Sung Ju Hwang

KAIST · DeepAuto.ai

Research Digest·12 November 2026·3 min read

Illustration: Linking memory back to raw data helps AI agents avoid repeating bad lessons

In one line

Evidence-linked memory, storing lessons with their source trajectories, reduces failures caused by incorrect or misapplied lessons in LLM agents.

Before you read this

Three ideas this paper builds on

agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.232 papers →evidence retrievalPulling specific, relevant information from a large corpus to support an answer, decision, or reasoning step.106 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.327 papers →

Read it as

Large language model agents can improve by storing lessons from past tasks, but these lessons are often wrong or misapplied, causing the agent to fail tasks it used to solve alone.

This paper introduces WikiTrace, a system that links each lesson to a summary of its original task and the full record of what happened. When an agent faces a new task, it can check the linked evidence to fill in missing details or correct the lesson before acting.

WikiTrace was tested on five benchmarks and improved performance by 10.7 points on average compared to an agent without memory, while cutting the rate of such failures from 6.5% to 4.5%. This shows that keeping evidence alongside lessons helps agents reuse experience more reliably.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agent memory evidence retrieval provenance verification failure attribution tool invocation