Are Online Skill and Memory Modules Always Worth Their Tokens? A Budget-Constrained Study of Web Agents

In three linesComparative study of memory and skill modules for web agents. On WebArena and WorkArena, a vanilla baseline with equivalent token budget matches or exceeds AWM, ASI, and ReasoningBank. Results across Gemini 3 Flash, GPT-4o-mini, Qwen 3.6-27B show apparent gains vanish against a budget-matched actor.

Summary generated by Claude — human-verified