Research
Research
Famyl's research focuses on representing human identity in a portable, model-agnostic form, maintaining long-term memory across years of conversation, and measuring personality fidelity — whether the people who knew someone recognise the companion.
Open problems we care about
- Identity representation that survives model migration
- Long-term memory that stays coherent over years, not sessions
- Personality fidelity metrics judged by people who knew the individual
- Grounding and refusal — knowing when a companion should say it does not know
- Efficient open-source voice and avatar models to make preservation affordable at scale
Our evaluation standard
The only benchmark that matters to us is recognition: does the family say 'that is exactly what she would have said'? Fluency without recognition is failure in this category.
Building in public
Famyl is early. We publish our thinking as we go and will share findings, limitations and negative results as the platform matures.
Frequently asked questions
- Can Famyl work with future AI models?
- Yes — identity is stored separately from the model that renders it, so companions can migrate to newer language, voice and video models.
- How do you measure whether a companion is accurate?
- By recognition from people who knew the person, rather than generic language benchmarks.