AlphaMemo: Self-Evolving AI Agent for Alpha Mining with Structured Memory
New LLM agent learns from failed factor searches to boost stock returns by 40%.
AlphaMemo, detailed in a June 2026 arXiv paper by Hang Yu, Zifan Zheng, Jeff Z. Pan, Tongliang Liu, Zhiyong Wang, and Fengxiang He, is a self-evolving LLM agent designed for financial alpha mining. Unlike prior systems that memorize only successful factors or entire search trajectories, AlphaMemo stores structured memory about which edit motifs worked or failed under specific parent-factor contexts. It extracts these motifs from Abstract Syntax Tree (AST) differences between successive factor expressions, applies a confidence-gated residual memory mechanism on top of a search-ledger prior, and uses asymmetric veto control to suppress high-confidence failure patterns. This allows the agent to avoid repeating past mistakes while efficiently exploring the combinatorial space of possible factors.
In experiments on Chinese CSI 500 and US S&P 500 datasets, AlphaMemo outperformed baselines in out-of-sample Sharpe ratios and fixed-budget discovery efficiency. Ablation studies confirmed the importance of each component: AST-diff-based motif extraction, residual learning with confidence gating, and veto memory all contributed to performance gains. The system addresses core challenges in AI-driven alpha generation: noisy, non-stationary feedback; redundant discoveries; and overfitting from naive reuse of past successes. Code is publicly available on GitHub.
- Uses AST-diff motif extraction from factor expression trees instead of full trajectories.
- Confidence-gated residual memory on a search-ledger prior prevents redundant discoveries.
- Achieved higher out-of-sample Sharpe ratios on CSI 500 and S&P 500 vs. baselines.
Why It Matters
AlphaMemo makes AI-driven quantitative finance more robust by learning from failure patterns, not just successes.