SplitMind-AI: Modeling LLM replies as competing internal pressures

Hi everyone,

I wanted to share SplitMind-AI, an open-source project exploring a different way to structure conversational LLM systems.

Instead of treating persona as a single prompt layer, SplitMind-AI models reply generation as a negotiation between competing internal pressures: desire, inhibition, defense, norms, and persona integration. The final response is generated from that tension rather than from tone alone.

The motivation is not psychological realism in a strict sense. It is inspectability. When a response feels off, I want to know whether the issue came from internal pressure, containment, safety constraints, or persona framing.

The current project includes:

• a Streamlit interface for chatting and inspecting traces
• explicit state for relationship, mood, drive, inhibition, and memory
• persistent vault-backed memory
• typed contracts between runtime nodes
• safety checks, output linting, and scenario-based evaluation scaffolding

It is still a research/architecture project rather than a polished end-user product, but I’d love feedback from people working on agent design, evals, and controllable generation.

Repo:

github.com

GitHub - yatarousan0227/SplitMind-AI: Psychodynamic-inspired AI agent architecture that...

Psychodynamic-inspired AI agent architecture that generates responses from structured internal tension instead of a single persona prompt.

Questions I’m especially interested in:

• Is this kind of decomposition actually helpful for controllability/debuggability?
• How would you evaluate “relational texture” or indirect emotional expression?
• Where would you draw the line between explicit rules and learned behavior?

Thanks for taking a look.