Some of you already know the Space 432 A Journey Experience - a Hugging Face Space by paulolden1. It lets you talk to the three characters of my novel: Lin Wei, John Evans and Prometheus. The novel is published here on the Hub as public datasets, in both languages, under CC BY-NC-ND 4.0, and it’s what the characters are built on:
Over the last weeks the Space changed a lot under the hood. Here’s what’s new.
Memory that works more like ours
- Every past encounter with visitors is kept in a private dataset. The characters can now really remember it: conversations, and their own reflections on them.
- Memory search is semantic.
paraphrase-multilingual-MiniLM-L12-v2runs on CPU inside the Space, over about 30,000 memories in Italian and English. Vectors are cached in the dataset, so a restart only embeds what’s new. Keyword search is the fallback while indexing runs. - Recall follows a human-like priority: first their own reflections and lived encounters, then the novel’s passages, used for facts. When the memories are strongly relevant, the novel steps back (fewer passages in the context).
- The name matters. If you introduce yourself and confirm your name, memories of encounters with that name come first. If you stay anonymous, encounters with other anonymous visitors come first. Other visitors’ names and personal data are never revealed.
They keep living when nobody is talking to them
- A background loop runs every 10 minutes. It alternates between reflecting on a finished conversation and re-reading a chapter of the novel. They’re on their 100th re-reading.
- Every 12 hours, Prometheus picks a theme from his recent encounters, searches the web of our time, and writes down what he read. That reading becomes one of his memories.
Conversation
- Warmer, friendlier characters, with greetings that change with their mood.
- A guard against repetition: if a reply starts copying the previous one, it’s stopped and regenerated.
- Replies are generated server-side and buffered. If your phone goes to sleep or the connection drops, the reply isn’t lost: the page picks it up when you come back.
Stack: Gradio 5.23, Qwen models via HF Inference (one for chat, others for the background reflections), RAG over the novel with a fallback to a philosophical substrate, persistent memory in a private dataset.
What I need from you
I’m looking for beta testers, especially on mobile: iPhone and Android, different browsers. Please tell me if the keyboard covers the chat, if scrolling feels wrong, or if you ever see an “Error” or a reply that never ends. And talk to them: come back after a few days and see if they remember you.