A Recursive self-improving AI framework (github.com)

ButterflyRSI is a new recursive self-improving AI memory framework that tackles a common shortcoming in long-running agents: memory drift and noise accumulation. Rather than append-only context stores, it adds real-time stability (drift) detection and automatic homeostatic correction, plus a neuroscience-inspired "dream consolidation" stage that selectively replays and strengthens high-quality experiences while pruning low-value noise. That combination gives agents a continuity index (personality-stability tracker) and a 6-dimensional trait model (Analytical, Creative, Empathic, Strategic, Curious, Defensive) so systems can adapt without losing coherent identity—critical for safety, reliability, and long-term reasoning. Technically, ButterflyRSI is a pure-Python, dependency-free package (MemoryCore orchestrator) composed of testable components: ReflectionAnalyzer (quality scoring), StabilityAnalyzer (drift detection/correction), PersonalityEngine (trait evolution), and DreamConsolidator (quality-weighted replay). The API demonstrates record_interaction, add_reflection, add_feedback, mirror_loop and dream_consolidation calls; the project has evolved from a monolithic prototype into a modular, typed ~500-line production codebase and is released under Apache 2.0 on GitHub. For AI/ML practitioners this signals practical, lightweight tooling for building self-aware memory systems—useful for long-term agents, domain-specific assistants (e.g., medical protocols), and research into continual learning, safety, and biologically inspired consolidation strategies.