Recursive Experiential Learning Monograph

Manuscript abstract

Abstract

Foundation-model agents are increasingly capable, yet many agentic workflows remain episodic: useful reasoning is discarded or retained only as raw history, failures are weakly propagated, and improvement is pursued through larger models or additional inference. We propose Continual Systems Intelligence (CSI), a system-level learning architecture that can adapt without task-specific updates to its foundation-model parameters. An Experience Compiler evaluates trajectories against independent evidence and proposes scoped episodes, claims, skills, deterministic tools, and context or routing policies. We call the conditional conversion of repeated verified reasoning into cheaper reusable computation Capability Compilation. Every promoted artifact remains provenance-linked and revocable: later contradictions, reverts, incidents, or real-world outcomes can invalidate its support. A trusted control plane separates execution from specification, authority, and verification. We specify a six-layer architecture spanning governed intent, isolated execution, experience and context, independent assurance, capability routing, and longitudinal evaluation. A prospective longitudinal software-engineering study will test whether accumulated verified experience reduces full lifecycle cost and human intervention per accepted outcome without degrading independent quality. CSI reframes continual agent improvement as a falsifiable question about the accumulation, verification, compilation, and governance of system-level experience.

• AI agents
• continual learning
• agent memory
• experience compilation
• human–AI collaboration
• evidence governance
• software engineering
• organizational learning

Attribution

Cite the manuscript

Cosmipher Research. (2026). Recursive Experiential Learning Monograph. Cosmipher Tech Labs. Manuscript.