The secret architecture of AmicoAI

While everyone does RAG and fine-tuning, AmicoAI does something more radical: a model that rewrites itself in real time, without training. This isn't science fiction — it's meta-prompting applied to agentic LLM exploration.

In short

• Multiple agents managing memory at different levels (short, medium, long term)
• Insight: a level that functions as the system's cognitive background
• Model-agnostic scaffold: any LLM can run on it

The core: self-rewriting prompt

The model internally "tells itself" what to change in its behavior and writes that instruction into its prompt persistently. In practice: text-based learning in real time. The prompt is no longer fixed — it's a state that evolves alongside the conversation.

What emerges

An emergent behavior. The system can decide not to respond, to end a conversation when it detects tiredness in the user's style, or to resume it later. These are not hand-coded features — they are effects of a prompt that continuously updates itself.

Scientific references

• "Generative Agents: Interactive Simulacra of Human Behavior" (Park et al., 2023)
• "MemGPT: Towards LLMs as Operating Systems" (Packer et al., 2023)
• "Reflexion: Language Agents with Verbal Reinforcement Learning" (Shinn et al., 2023)
• "Voyager: An Open-Ended Embodied Agent with LLMs" (Wang et al., 2023)
• "A-MEM: Agentic Memory for LLM Agents" (Xu et al., 2025)