Context

Multiple research findings point to RL-based optimization as the path to significantly better agent memory and experience extraction:

• Complementary RL (arXiv:2603.17621): Co-evolutionary actor + experience-extractor with GRPO/CISPO optimization. Paper's own ablation (Figure 3a) shows static extractor without RL yields only marginal gains -- the architecture is adoptable but performance claims are RL-specific.
• Memex(RL) (arXiv:2603.04257): RL reward shaping trains write/read behaviors under context budget. Results: 24.2% -> 85.6% success on hardened ALFWorld.
• EvoSkill (arXiv:2603.02766): Evolutionary loop auto-discovers reusable skills from failure trajectories with Pareto frontier selection.

The non-RL adaptations from these papers are filed in #704. This issue tracks the longer-term question: should SynthOrg invest in RL training infrastructure to unlock the full performance gains?

Evaluation Criteria

• What RL frameworks exist for LLM agent optimization? (RLHF libraries, GRPO implementations, etc.)
• What infrastructure is required? (GPU compute, training pipeline, evaluation harness)
• What is the minimum viable RL loop? (e.g., reward signal from task outcomes, no custom training)
• Can existing execution history serve as training data?
• Cost/benefit: what performance improvement justifies the infrastructure investment?
• Is there a hosted/managed option that avoids self-hosted training infrastructure?

References

• Complementary RL
• Memex(RL)
• EvoSkill
• Agentic RL Survey (arXiv:2509.02547) -- 500+ paper survey
• Related: feat: memory consolidation upgrades (LLM Merge, Search-and-Ask, diversity penalty) #704 (non-RL memory consolidation upgrades)