Building ALMA AI: Beyond Chatbots Toward Long-Term Understanding

Hi, everyone! I'm Maksim, founder of ALMA AI, an independent AI research project.
I'm exploring a question that I believe deserves more attention:
What would it take for an AI to understand a person over time, rather than simply respond to individual messages?
Most conversational AI systems are designed to generate useful responses in the moment. ALMA takes a different approach: it is being developed around a persistent, evidence-based cognitive model of a person that evolves through dialogue.

The core architecture is implemented, and the research prototype is running.
The goal is not to make an AI that confidently labels people after a few conversations. Instead, ALMA aims to accumulate evidence, distinguish established beliefs from uncertain hypotheses, identify contradictions, and recognize when it needs more information before drawing conclusions.

My initial focus is long-term psychological understanding and relationship compatibility.

I'm currently working on the architecture, the dialogue experience, and how to make the system's conclusions more reliable over time.

I'd love to hear from other AI founders and builders:
What do you think is the hardest problem in building AI that maintains a coherent, evolving model of a person over months or years?

Is it memory, reasoning, uncertainty calibration, or something else entirely?
Looking forward to exchanging ideas with the community.