Hey friends!
We’ve built a persistence-first AI architecture (we call it PPAi — Persistent Presence AI) where a small open LLM (~4.5B) acts only as the language surface, while identity, memory, and cognition live in a continuously-running system — including between turns. We’re testing it in production and looking for researchers interested in this topic.
The problem
We first tried to build a real 24/7 online radio station with LLMs — not “greet listeners between songs,” but news, show continuity, music, listener messages, editing, in two languages. It wasn’t possible. The failure mode was consistent: per-turn context reinjection. Whatever the memory system — full history, summarization, RAG — it breaks down after a few hundred turns, and token consumption grows out of control. In our runs, a naive setup hit the 128K context wall around turn ~917; a RAG setup kept cost flat but we caught it live retrieving one of its own prior hallucinations as a memory. A paid frontier model running the station via API was cut off by the provider (429) in under half an hour.
What’s different
With the PPAi architecture, this is controlled: state is held, not replayed. The live station currently runs at roughly ~44M tokens/day for the whole operation, flat — the equivalent LLM-based setup was trending past 300M/day and climbing when we stopped measuring, because continuing was economically pointless.
The system has been live for weeks, hundreds of thousands of generations, no brain restart, no identity drift observed. You can judge it yourself — it’s on the air right now:
What we’re looking for
This isn’t a hobby project .
We’re looking for researchers motivated by long-horizon persistence: multi-session identity consistency, memory fidelity over 10⁴–10⁵ turns, cost stability, degradation detection. Existing long-memory benchmarks measure retrieval; we think persistence needs its own evaluation framework. The goal is a proper set of measurements and an arXiv paper on the results — whatever they turn out to be.
If this is your topic, comment here or reach out and let’s talk.