OpenClaw Auto-Dream adds memory consolidation for agents

This is less a cute metaphor than a real architectural bet: if agents run continuously, memory maintenance needs to be automatic, periodic, and opinionated.

• –The core loop scans recent logs, extracts decisions and facts, dedupes them, and routes them into five memory layers, which is a lot more robust than a single `MEMORY.md`.
• –The importance scoring and forgetting curves are the interesting part: they acknowledge that not all memories should survive equally, which is exactly where most agent memory systems fall apart.
• –The dashboard, health metrics, and export/import flow suggest this is meant for serious 24/7 agents, not one-off demos.
• –The tradeoff is complexity: you are adding a second system that needs tuning, monitoring, and trust before it starts paying back the overhead.
• –Because it is tightly tied to OpenClaw/MyClaw, the audience is narrow, but the idea maps cleanly to any agent stack that needs long-horizon continuity.