Corvus: Successful but Inapplicable Memories in Tool-Using Support Agents
Independent Research · 2026
Independent research on agent memory and evaluation. I tested whether past trajectories that were successful in their original situation can become misleading precedents when a similar current task requires a different decision.
Across 100 synthetic tasks, five memory conditions, and two GPT-4.1 variants, I observed no average critical-decision penalty from a misaligned similar precedent relative to an action-matched control. The project includes a reproducible evaluation harness, paper, and interactive trace explorer.