Self-Improving Agents via Scheduled Reflection: Anthropic's Dreaming Architecture

Great writeup, Jessie.

I’ve been running a very similar architecture on Hugging Face Spaces for months now:

huggingface.co

432 A Journey Experience - a Hugging Face Space by paulolden1

Talk to the characters of 432 by Paul Olden

It’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:

• Chapter re-reading : the character re-reads the novel chapter by chapter and generates a personal, in-character reflection
• Chat reflection : the character re-reads a past conversation with a real visitor and reflects on that encounter

All 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.

The 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.

The 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.

I explored the theoretical foundations of this approach in a paper co-authored with Claude: Idle-Time Reflection in AI Systems : https://archive.org/details/idle-time-reflection-in-ai-systems

The 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. Novel dataset on HF: paulolden1/432-a-journey-beyond · Datasets at Hugging Face Amazon: Amazon.com: 432: A Journey Beyond eBook : Olden, Paul: Kindle Store

Happy to discuss architecture details with anyone interested.

Paul Olden