← Key papers

Paper review · November 12, 2026

Don't trust memory blindly: Reducing memory-induced performance degradation in LLM agents with grounded memory

Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

AgentsLanguage models Intermediate Upvotes 45 cs.AI KAIST AI

By attaching to each lesson (insight) a link to the original trajectory so the agent can verify it directly, the method improved memory performance while reducing failures caused by memory.

Problem

External memory compresses trajectories from past tasks into short lessons for reuse. But this causes "memory-induced degradation," where the agent fails tasks it could solve without memory because of the memory. Lessons are inferred from a single success/failure label, so they can be wrong. Yet they are provided without the original trajectory, leaving the agent no way to verify them. The analysis found that 52% of degradation cases stemmed from memory entries the agent followed.

Method

WikiTrace stores memory in three linked layers. From the top: short insights, compilations summarizing why a run succeeded or failed, and the original trajectories. On a new task, the agent starts from all insights. If an insight lacks execution details, conflicts with another insight or came from a failed run, the agent follows the links down using terminal tools (read, bash). There it fills in missing information or corrects the insight before acting.

Results

Limitations

Memory is built once from a fixed corpus of past runs and not updated afterward. The main experiments used only deepseek-v4-flash (the gains are reported to hold on other backbones as well). Because the agent decides whether to descend to the original, the authors acknowledge that the Descent vs. Stay comparison is hard to interpret as a causal effect. The classification of failure causes relies on an LLM judge.

Why it matters

The work shows that when adopting agent memory, one should look not only at average accuracy but separately at whether it "breaks what used to work." A simple design that attaches sources to lessons can improve reliability, making it directly useful for designing experience reuse in practical agents.

Key terms

Memory-induced degradation
A phenomenon in which an agent fails a task it would have succeeded at without memory, because of memory that is wrong or wrongly applied.
Gain / Regression
Gain is the rate at which memory turns failures into successes; Regression is the rate at which it turns successes into failures. This separates two effects hidden by average accuracy.
Trajectory
A sequential record of the actions and observations an agent produced while performing a task.

Original

Authors: Yukyeong Lee, Yumin Choi, Jinheon Baek, Sung Ju Hwang · Posted 2026-11-10 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.