AnnenkovLabs/girl-agent

What it solves

It addresses the lack of realism in AI personas, moving beyond simple system prompts to create a Telegram-based AI companion that behaves like a human. It solves the problem of "AI-like" responses (e.g., overly polite, always available, repetitive phrasing) by implementing a state-driven behavioral engine.

How it works

Instead of relying on a single prompt, the agent's behavior is constructed from multiple layers of state:

• Presence & Sleep: Simulates human availability, including sleep cycles and varying response frequencies.
• Daily Life: Follows a schedule (work, study, free time) that affects accessibility.
• Relationship System: Tracks five relationship metrics (interest, trust, sympathy, irritation, cringe tolerance) and progresses through nine stages of intimacy.
• Conflict System: Triggers silence or withdrawal if the user spams or pushes boundaries.
• Memory: Stores significant events in a long-term markdown file for future recall.
• Anti-AI Layer: Specifically prohibits typical LLM markers like markdown formatting, generic phrases ("I understand"), and emoji strings.
• Userbot Mode: Uses MTProto to act as a real Telegram account, allowing it to read, react, type, and edit messages.
• Agenda: Proactively plans messages (e.g., wishing luck for an event) to simulate an independent life.

Who it’s for

Developers and enthusiasts looking to create highly realistic, stateful AI personas for Telegram that avoid the typical patterns of a chatbot.

Highlights

• State-driven behavior: Uses a complex system of counters and stages rather than just a system prompt.
• Longe-term memory: Persists important events to maintain continuity over months.
• Human-like interaction: Simulates typing, reactions, and strategic silence.
• Proactive engagement: The agent can initiate conversations based on its own internal agenda.
• Deployment flexibility: Supports local installation, Docker, and a native Rust desktop client.