Study: Giving AI Agents Memory of Past Work Mostly Makes Them Worse
alex_verem · x · 2026-08-26
A study from Stony Brook University tested 11 models across three benchmarks and found that giving agents a "skill memory" often degrades performance. Key findings:
- Task-level memory drags performance down: Saving skills from whole tasks (e.g., "buy all weightlifting benches") lowered success rates below the no-memory baseline by up to 7.4 points.
- Granularity is the fix: Breaking tasks into steps and saving subtask skills (e.g., log in, check cart) improved performance.
- Text beats code: Skills stored as plain text transferred better than those stored as Python functions, as text adapts better while functions carry hardcoded assumptions.
The result challenges the common practice of saving full workflows for agent memory.
More from coding & agent
- AgentConnect: Open-Source Alternative to Claude Tag for Multi-Agent Workflows — Div_pradeep · 2026-08-26
- Why Cloud Agents Are the Future: Settling on Conductor and Replicas — jarrodwatts · 2026-08-26
- LangChain open-sources WikiBench to measure how much codebase wikis help coding agents — LangChain · 2026-08-26
- GenOS Multi-Agent Orchestration Framework Seeks Extreme Stress Tests — MonokoEloba · 2026-08-26
- The Downside of AI Coding: Blowing Past the MVP Without Stopping — talkaboutdesign · 2026-08-26
- Perplexity Brain lets agents explore user history on demand — perplexity_ai · 2026-08-26