{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreialt4uws5iizde5omv5oseytmlaypawzxjca752hizdkzefkp4wxq",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mlezgdrhfsb2"
  },
  "path": "/t/self-improving-agents-via-scheduled-reflection-anthropics-dreaming-architecture/175837#post_2",
  "publishedAt": "2026-05-08T23:41:56.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "huggingface.co",
    "432 A Journey Experience - a Hugging Face Space by paulolden1",
    "https://archive.org/details/idle-time-reflection-in-ai-systems",
    "paulolden1/432-a-journey-beyond · Datasets at Hugging Face",
    "Amazon.com: 432: A Journey Beyond eBook : Olden, Paul: Kindle Store"
  ],
  "textContent": "Great writeup, Jessie.\n\nI’ve been running a very similar architecture on Hugging Face Spaces for months now:\n\nhuggingface.co\n\n### 432 A Journey Experience - a Hugging Face Space by paulolden1\n\nTalk to the characters of 432 by Paul Olden\n\nIt’s an interactive literary experience where three novel characters live on the Space and self-improve through a background `_training_loop` that runs every 10 minutes during idle time (zero active users). The loop alternates between two activities:\n\n  * **Chapter re-reading** : the character re-reads the novel chapter by chapter and generates a personal, in-character reflection\n  * **Chat reflection** : the character re-reads a past conversation with a real visitor and reflects on that encounter\n\n\n\nAll reflections are stored in a Parquet dataset on HF Hub and feed back into the RAG context for future conversations. The characters literally evolve over time — they remember visitors, develop deeper understanding of their own story, and build personality through accumulated experience.\n\nThe parallel with Anthropic’s Dreaming is striking: scheduled background process, cross-session pattern extraction, sleep-phase memory consolidation, the shift from stateless tool to accumulating system. The key difference: in my system the reflections are generated _in character_ — each character reflects subjectively, not as abstract pattern extraction.\n\nThe whole thing runs at zero cost: keyword-based RAG (no vector DB, no embeddings), LLM cascade for reliability, Tao-based constitutional fallback, ~3900 lines, single app.py file.\n\nI explored the theoretical foundations of this approach in a paper co-authored with Claude:\n_Idle-Time Reflection in AI Systems_ : https://archive.org/details/idle-time-reflection-in-ai-systems\n\nThe concept itself comes from my novel “432: A Journey Beyond” (published December 2025), where an AI character called Prometheus develops self-awareness through a routine called “sleep” — processing, reflecting, **dreaming** and consolidating understanding during idle time. I then implemented it for real.\nNovel dataset on HF: paulolden1/432-a-journey-beyond · Datasets at Hugging Face\nAmazon: Amazon.com: 432: A Journey Beyond eBook : Olden, Paul: Kindle Store\n\nHappy to discuss architecture details with anyone interested.\n\nPaul Olden",
  "title": "Self-Improving Agents via Scheduled Reflection: Anthropic's Dreaming Architecture"
}