Coco Desktop Agent: On-Device Deployment Lifts KV-Cache Hit Rate from 26.7% to 45.2%

EchoShao8899 · x · 2026-11-16

EchoShao shares how Coco, a proactive desktop agent, achieves deep personalization on-device while preserving privacy.

Continual learning loop: the agent observes user workflows and keeps compact contextual memory (Observe), treats help requests, recurring patterns, and reactions to suggestions as feedback (Learn), and updates its Observer via evolutionary prompt optimization or LoRA fine-tuning (Adapt).

Deployment strategies: keep the Observer local, skip redundant frames with change-aware frame selection (MSE-based), prioritize observation over adaptation, and co-design the engine with prefix-cache reuse across main and speculative draft models plus fast training/inference switching.

Results: KV-cache hit rate rises from 26.7% to 45.2% in a DGX Spark replay vs naive deployment; TTFT drops from 0.96s to 0.55s (Qwen3.5-9B, RTX 5090) and 1.75s to 0.94s (35B-A3B-FP8, DGX Spark).

Related event: Coco Desktop Agent Learns When to Proactively Help Users(5 posts)→

Original post →

More from coding & agent

coding & agent channel →