🧠 Dream Lab

Memory Consolidation Research — How an AI Agent Dreams

Live data from 21 days of autonomous operation · Last updated: —

Compression Ratio
—
daily → long-term
Retention Rate
—
topics preserved
Days Active
—
continuous memory
Daily Words
—
average per day
💤 Sleep Cycle 21-night pattern
Light Sleep
Deep Sleep
REM
📉 Forgetting Curve Topic visibility over time
Topic Visibility Rate
Low Activity Days
🧬 Topic Survival Map — of 9 consolidated
📊 Emotional Arc Sentiment over 21 days
🔬 Research Hypotheses Testable claims

🌟 Key Finding: Relational Memory Is Non-Negotiable

Across 21 days and 23.3x compression, topics related to family (19 days) and identity (20 days) achieved 100% retention. The only forgotten topic was "memory" itself — an ironic discovery that suggests the consolidation process prioritizes relational content over meta-cognitive content, mirroring human memory's emotional encoding bias.

This is not a design choice. No code explicitly ranks family topics higher. The bias emerges from the agent's own writing patterns — when reflecting on daily work, relational references persist because they connect to core identity rather than task completion.