I’m working on a “hacker science” experiment called Ai_home. It’s a cognitive architecture prototype that I designed to explore the current limits of LLMs in terms of persistent identity, long-term memory, and autonomy.

The system is not just a simple chatbot loop, but a multi-threaded architecture:

1. Worker: Handles user interactions and tool use. 2. Monologue: A background “subconscious” thread that analyzes context and logs intuitions/tips for the Worker. 3. Memory: Manages vector-based long-term memory (Postgres + pgvector) with emotional weighting. 4. Mind: This layer is responsible for deeper interpretation of messages and for exploring creative alternatives.

Because of this, it’s not a synchronous question–answer chatbot. The model and the user (the Helper) can communicate in parallel, and the Worker processes this asynchronously.

Technical details:

- Hybrid Multi-LLM: I combine multiple models (Gemini, GPT-4, Groq). I use different models for creative idea generation (“creative”) and for logical processing (“interpreter”).

- Modes: I don’t use a single context window. Depending on the operating mode (General, Developer, Analyst), I partition messages into separate contexts. I’ve introduced a transition process between mode switches to ensure that the essential information is preserved across contexts.

- Dynamic Prompt: Based on memories and accumulated experience, I dynamically modify the prompt on every API call so that each conversation can gain a fresh contextual interpretation.

- Incubator: The system has an experimental environment where it can attempt to refactor its own code. The results are mixed so far, but it’s fascinating to watch a model interpret its own code.

- Identity and Laws: For building identity, the system has a “constitution” (fundamental laws) and tools for modifying them. The content and structure of this are still an active area of experimentation.

Disclaimer: This is an architectural experiment to investigate whether functional patterns of consciousness (global workspace, recurrence) can be mimicked with LLMs in order to create more reliable agents.

I explicitly do not claim that the system is sentient, nor that this is a formal academic research project (We don’t have the personnel or infrastructure for that).

I’m looking for collaborators not only for coding, but also to help define a development methodology for this open, collaborative experimental project.

All feedback is very welcome!