Hours-long conversations without forgetting

Built for All-Day Embodied Scenarios

A reception robot at an exhibition runs 8+ hours. A companion robot at a care facility talks all day. Previously, after ~30 exchanges the AI's context window filled up and behavior degraded — it forgot instructions, repeated itself, or lost track of the tour.

What's New

• Automatic context compaction — Older exchanges are summarized and compressed, keeping the context window fresh
• Important information is preserved — System instructions, user preferences, and current task state are never dropped
• Smooth degradation — Instead of a hard cliff, quality stays consistent across hundreds of turns
• Optimized for embodied patterns — Tool call results (sensor data, navigation confirmations) are aggressively compressed since their value is short-lived

Real-World Impact

• Guided tours: A 20-station tour with explanations at each stop now completes without the robot forgetting where it is
• Reception: All-day greeting duty without personality drift or instruction loss
• Companionship: Extended conversations remain coherent and contextual

No configuration needed — this works automatically on all robots.