The Trap of Trajectory: Towards Understanding and Mitigating Spurious Correlations in Agentic Memory

TL;DR

This paper investigates the vulnerability of agentic memory in LLMs to spurious correlations, benchmarks their impact, and introduces CAMEL, a calibration method that mitigates these issues while maintaining performance.

Contribution

It provides a benchmark for spurious patterns in agentic memory and proposes CAMEL, a calibration technique to reduce reliance on spurious correlations in memory systems.

Findings

Memory improves reasoning on clean inputs.

Memory reliance on spurious patterns increases with their presence.

CAMEL reduces reliance on spurious patterns across architectures.

Abstract

Agentic memory enables LLMs to persist information beyond a single context window and reuse it in later decisions, but it also introduces a new vulnerability: spurious correlations, where retrieved memory carries miscorrelated evidence and propagates erroneous reasoning into downstream decisions. Despite the widespread use of agentic memory, this risk remains largely underexplored. We address it from two aspects. First, we benchmark several canonical types of spurious patterns identified through causal structure and record them across trajectory-level memory. Diagnosing agentic memory systems on this benchmark reveals that memory improves reasoning on clean inputs but amplifies reliance on spurious patterns when they are present. Second, we propose CAMEL, a plug-and-play calibration method that operates across diverse memory architectures at both write and retrieval time. CAMEL consistently reduces reliance on spurious patterns across all three types while preserving or improving performance on clean inputs and staying robust under adaptive attacks targeting the calibration. Overall, CAMEL offers a principled and lightweight solution toward more reliable agentic memory deployment.