DeepSeek Harness - MiMoCode Workspace Auto-Memory Plugin

DeepSeek Harness - MiMoCode Workspace Auto-Memory Plugin aims to address the difficulty agents have in maintaining long-term memory under limited context windows. Built on the MiMoCode architecture, the plugin achieves cross-session persistence through a four-layer hierarchical memory mechanism: automatically injecting core facts and architectural contracts at session start (Layer 1), recording task checkpoints (Layer 2), asynchronously distilling and consolidating new knowledge in the background (Layer 3), and providing an explicit disk-persistence tool (Layer 4). The plugin supports both global activation and workspace isolation, enabling the agent to understand the project context from the outset without manual retrieval. It uses `MEMORY.md` and `checkpoint.md` to manage knowledge and silently updates during idle periods, without consuming conversational tokens or increasing response latency, making it highly suitable.