self-evolving-agent-experience

name: self-evolving-agent-experience category: research created: "2026-05-19" source: "arXiv:2605.15461v1" description: Framework for LLM agents that accumulates and reuses cross-task experience including verified skills, statistical evidence of effective strategies, and recurring error-fix patterns. Enables zero-test-time search on new tasks. tags: [agent, memory, self-evolving, experience-reuse, drug-discovery]

Self-evolving Agent Experience (DrugSAGE)

Source: arXiv:2605.15461v1 - "DrugSAGE: Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery"

Summary

Proposes a framework for LLM-based agents that accumulates cross-task memory of verified skills, statistical evidence about effective strategies, and recurring error-fix patterns. Eliminates redundant trial-and-error search on new tasks, achieving SOTA performance without test-time search after sufficient experience accumulation.

Core Methodology

Key Problem

Current LLM agents can find SOTA solutions through extensive trial-and-error but do not retain experience across tasks, paying the full search cost on every new task.

Experience Components

1. Verified Skills: Solutions that worked on previous tasks
2. Statistical Evidence: Data about which strategies are most effective
3. Error-Fix Records: Recurring errors and their fixes

Operating Modes

1. Direct Transfer: Apply a working solution from memory without any test-time search
2. Guided Search: Use accumulated experience to inform and narrow the search space

Results

• 33 molecular property prediction tasks: ranks 1st among 9 SOTA agents in single-task setting
• Cross-task: 0.935 normalized score on 17 held-out tasks after learning from 16 smaller tasks
• Outperforms all baselines by 10-30% in zero-test-time search regime

When to Use

• LLM agents performing repetitive task families (drug discovery, ML engineering, etc.)
• Scenarios where agents repeatedly rediscover the same solutions
• Any domain where experience accumulation can eliminate redundant search

Implementation Considerations

• Requires structured memory across task boundaries
• Need mechanisms for verifying and indexing successful strategies
• Error-fix patterns should be generalized, not task-specific
• Can generalize beyond drug discovery to any agent-based workflow

Activation

self-evolving agent, cross-task memory, experience reuse, agent skill accumulation, drugSAGE