Agent Memory Decay Under Controlled Web Drift

• Duration: Jan 2026 - Present
• Context: EECS 545 (Machine Learning) Course Project
• Advisor: Prof. Honglak Lee
• Link: Progress Report

Project Overview

While agentic memory is a rapidly growing area, existing methods primarily focus on static environments or cross-task generalization. This project investigates a critical unanswered question: how does test-time agent memory (e.g., semantic insights vs. procedural workflows) degrade when the underlying website environment evolves?

Key Contributions

• Developed a fully reproducible testbed using Docker, injecting 6 controlled drift types (surface, structural, access, content, process, runtime) into a self-hosted open-source web application.
• Extracted and implemented different abstraction levels of agent memory, comparing semantic insights (ExpeL) with procedural workflows (AWM) injected into LLM contexts.
• Formulated an evaluation framework measuring Experience Transfer Gap (ETG) and Experience Decay Rate (EDR).
• Evaluated agent robustness under varying drift settings, uncovering critical empirical correlations between environment coupling and memory vulnerability.