AlphaMemo: Structured Search-Process Memory for Self-Evolving Alpha Mining Agents

AlphaMemo is a new LLM-based agent framework that improves automated financial factor discovery by using structured memory of past search patterns rather than naive trajectory replay. The system records reusable evidence about which code modifications succeed or fail in specific contexts, demonstrating better out-of-sample performance on major indices while reducing redundant exploration.

AlphaMemo addresses a fundamental problem in machine learning-driven quantitative finance: how to efficiently explore the massive space of possible trading factors without overfitting to historical data. Traditional approaches either memorize complete solution paths—leading to overfitting—or ignore past attempts entirely, forcing agents to rediscover failed strategies repeatedly. This research bridges that gap by extracting Abstract Syntax Tree differences between factors to identify reusable "edit motifs" that work or fail under specific conditions.

The broader context reflects growing interest in using language models for systematic alpha discovery. Financial institutions increasingly recognize that LLMs can combine domain knowledge with symbolic reasoning to generate novel factors, but previous implementations struggled with search efficiency and generalization. AlphaMemo's innovation—confidence-gated residual memory paired with asymmetric veto control—allows agents to learn which modifications are likely to fail before wasting computation.

For the quantitative trading industry, this matters considerably. Enhanced factor discovery efficiency reduces computational costs while improving risk-adjusted returns. The experimental validation on CSI 500 and S&P 500 datasets demonstrates real-world applicability across different markets, suggesting the approach scales beyond academic benchmarks.

The open-source release signals that this methodology will likely influence how quant shops architect their automated research pipelines. Success here could accelerate the timeline for deploying AI agents in production alpha generation, though the persistent challenge of regime change and market adaptation remains unsolved.

• →AlphaMemo uses structured memory of code edit patterns rather than full trajectories, reducing redundant exploration and overfitting in factor discovery
• →The system extracts reusable evidence from Abstract Syntax Tree differences to identify which modifications work in specific parent-factor contexts
• →Experimental results show improved out-of-sample performance on major indices compared to baseline approaches
• →Confidence-gated residual learning and asymmetric veto control enable the system to suppress high-confidence failure patterns efficiently
• →Open-source release indicates this approach will likely influence production quantitative trading research infrastructure