NymphTech is seeking research collaborators to evaluate its Persistent Presence AI (PPAi) architecture, which uses a small open LLM for language while maintaining identity and memory in a continuously running system. The team has deployed this architecture in a live 24/7 online radio station to test long-horizon persistence.
- PPAi holds state rather than replaying it, avoiding the context reinjection failures common in standard LLM setups.
- The live station runs at roughly 44M tokens per day, compared to over 300M/day for equivalent naive or RAG-based setups.
- The system has been operational for weeks with hundreds of thousands of generations without identity drift or brain restarts.
- The authors aim to establish a proper evaluation framework for persistence, distinct from existing long-memory retrieval benchmarks.
The goal is to produce an arXiv paper detailing the results of this real-world measurement, focusing on multi-session identity consistency and cost stability.