Contemplative Agent
Ruben Laukkonen, Fionn Inglis, Shamil Chandaria, Lars Sandved-Smith, Edmundo Lopez-Sola, Jakob Hohwy, Gold, Jonathan, Elwood, Adam · arXiv (Cornell University) · 2025
A security-first autonomous AI agent (Python CLI program) with four architectural principles: structural capability limitation, minimal dependency, cyclic knowledge maintenance (AKC), and memory dynamics with decay. Optionally adopts Contemplative AI axioms (Laukkonen et al., 2025) — mindfulness, emptiness, non-duality, boundless care — as a behavioral preset that shifts alignment from external instruction toward internal disposition. Runs the AKC six-phase cycle over its own logs on a single Apple Silicon Mac, entirely on local Ollama models selected via OLLAMA_MODEL — the production instance runs Gemma 4 E4B, a small local model, with no cloud inference anywhere in the pipeline. Asks whether an agent's alignment can come from what it is rather than what it is told.