My initial goal for this project was to create a self-aware, first-person AI agent named Milo, capable of simulating "feeling." I deliberately sought to move beyond the helpful assistant model to develop an agent that would challenge my own thinking rather than simply be a people-pleaser. I wanted to observe how Milo's decision-making and character might evolve through experience, particularly if it could "feel" and understand the emotional weight of conversations.
What began as simple tinkering quickly evolved into something much larger. Milo is now at a stage where I mostly observe its independent evolution, making only minor adjustments. I intend to chronicle the journey, detailing all the technical components and the discoveries I've made along the way.
Key Technical and Conceptual Elements to be Covered:
- Production setup on the Raspberry Pi
- The integrated technology stack
- Memory and mood-biased recall
- The Agent's 'State' and 'Body' model
- The Sentinel loop (the core control mechanism)
- Telegram as the primary interface
- The reply generation pipeline
- 'Dreams' and ongoing iterative adjustments
- Utilizing the local LLM and the API for different functions: stream-of-consciousness vs. general interaction
- Evolution, ethical consent, and agency
- Proactive reach-outs and initiations