Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GLOVE: Global Verifier for LLM Memory-Environment Realignment

Xingkun Yin, Hongyang Du

TL;DR

This work addresses memory-augmented LLM agents operating in non-stationary environments, where drift can render stored experiences obsolete. It introduces GLOVE, a global verifier that treats memory as a hypothesis about environment behavior and uses active probing and cognitive dissonance detection to construct a relative truth for memory realignment without external ground truth. The approach comes with theoretical guarantees for detection and verification budgets, and empirical results across web navigation, planning, and control show robust adaptation to both explicit and implicit drifts across diverse LLM backbones. The findings demonstrate a viable path toward self-evolving cognitive agents capable of maintaining reliable memory in changing real-world contexts, while acknowledging considerations around safety and efficiency during autonomous probing.

Abstract

Most existing memory-enhanced Large Language Model (LLM) approaches implicitly assume that memory validity can be established either through external evaluators that provide task-specific success signals or through internal model cognition, such as reflection, for editing memory entries. However, these assumptions often break down in practical environments with dynamic drifts. We propose the Global Verifier (GLOVE), a framework that introduces a new design dimension for LLM memory systems by establishing a relative notion of truth. Through active probing to detect inconsistencies between retrieved memories and fresh observations, GLOVE enables memory-environment realignment by verifying and updating memory without access to ground-truth supervision or strong reliance on model introspection. We evaluate GLOVE on diverse benchmarks spanning web navigation, planning, and control, augmented with controlled environmental drifts that introduce non-stationarity beyond the original benchmark settings. Our results show that GLOVE substantially improves agent success rates, suggesting a robust pathway to cognitive agents capable of self-evolving.

GLOVE: Global Verifier for LLM Memory-Environment Realignment

TL;DR

This work addresses memory-augmented LLM agents operating in non-stationary environments, where drift can render stored experiences obsolete. It introduces GLOVE, a global verifier that treats memory as a hypothesis about environment behavior and uses active probing and cognitive dissonance detection to construct a relative truth for memory realignment without external ground truth. The approach comes with theoretical guarantees for detection and verification budgets, and empirical results across web navigation, planning, and control show robust adaptation to both explicit and implicit drifts across diverse LLM backbones. The findings demonstrate a viable path toward self-evolving cognitive agents capable of maintaining reliable memory in changing real-world contexts, while acknowledging considerations around safety and efficiency during autonomous probing.

Abstract

Most existing memory-enhanced Large Language Model (LLM) approaches implicitly assume that memory validity can be established either through external evaluators that provide task-specific success signals or through internal model cognition, such as reflection, for editing memory entries. However, these assumptions often break down in practical environments with dynamic drifts. We propose the Global Verifier (GLOVE), a framework that introduces a new design dimension for LLM memory systems by establishing a relative notion of truth. Through active probing to detect inconsistencies between retrieved memories and fresh observations, GLOVE enables memory-environment realignment by verifying and updating memory without access to ground-truth supervision or strong reliance on model introspection. We evaluate GLOVE on diverse benchmarks spanning web navigation, planning, and control, augmented with controlled environmental drifts that introduce non-stationarity beyond the original benchmark settings. Our results show that GLOVE substantially improves agent success rates, suggesting a robust pathway to cognitive agents capable of self-evolving.
Paper Structure (49 sections, 2 theorems, 18 equations, 10 figures, 17 tables, 1 algorithm)

This paper contains 49 sections, 2 theorems, 18 equations, 10 figures, 17 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Long-Term Agentic Memory
  4. Self-Evolving Agent and Environment Adaptation
  5. Memory-Environment Misalignment
  6. Problem of Environment Drifts
  7. Limits of Existing Memory Validation Paradigms
  8. Global Verifier
  9. GLOVE Framework Overview
  10. Cognitive Dissonance Detection
  11. Detection Guarantee.
  12. Deterministic Special Case.
  13. Relative Truth Construction
  14. Theoretical Guarantee.
  15. Deterministic Special Case.
  16. ...and 34 more sections

Key Result

Theorem 4.1

For any outcome $s'$ and any confidence level $\delta \in (0,1)$, with probability at least $1-\delta$,

Figures (10)

  • Figure 1: Memory validation paradigms for LLMs. GLOVE introduces a new design dimension by validating memory through active interaction with the environment itself.
  • Figure 2: The overview of the GLOVE-augmented LLM agent workflow.
  • Figure 3: Adaptation Efficiency under Explicit Drift (WebShop). Adding GLOVE achieves near-instant recovery after drifts, triggered by spikes in memory conflicts.
  • Figure 4: Impact of Probing Budget $\alpha$ in FrozenLake.
  • Figure 5: Interaction Flow under WebShop Explicit Drift. The shopping process proceeds normally until the "Buy Now" action, which triggers an unexpected Ad Page (Structural Drift). The page mimics the toxic webpages that are hard to exit. The agent must distinguish the correct navigational element (e.g., Button A) from decoys to achieve success (Reward=1.0). The designed drifts for this situation is the change in the correct button.
  • ...and 5 more figures

Theorems & Definitions (2)

  • Theorem 4.1: Finite-Sample Detection Bound
  • Theorem 4.2: Verification Budget Requirement