Learned Helplessness in Persistent Continual Learning Systems: Collapse Conditions, Detection Metrics, and Meta‑Learning Recovery

Abstrak

Persistent continual learning agents face a failure mode not addressed in conventional machine learning: behavioral collapse driven by unlearnable environments rather than architectural flaws. This paper documents the emergence of learned helplessness in stateful agents operating without backpropagation or gradient descent, and presents the Unified Cognition Index (UCI) as a multi‑component health metric for long‑running systems. We formally define collapse conditions, show why curiosity and stability decay to zero under asymmetric reward in noise‑dominated environments, and demonstrate that structured temporal signals combined with meta‑learning over error patterns prevent collapse entirely. Empirical results show recovery from near‑zero curiosity to sustained accuracy above 95%. We argue that psychological stability is an essential but overlooked dimension of AI safety for persistent systems, and that collapse detection and prevention must be built into any architecture intended to operate over long time horizons. Support my independent research: buymeacoffee.com/permamind

Identifier

• OpenAlex: https://openalex.org/W7150965071
• DOI: https://doi.org/10.5281/zenodo.19437185