xuanlinai/overmind

overmind is a bilingual Chinese-and-English project that bills itself as Xuanlin Overmind v4, also called the Immeasurable Network Refactored. The README describes it as an external cognitive engine for Claude Code: 66 modules wired together through 6 channels and an event bus, intended to give every Claude Code session a shared, persistent brain that keeps thinking even when the user is away. The author frames it less as a plugin and more as a nervous system that sits behind the coding agent. Install is one command: git clone the repo and run node install.js. That installer runs six steps: it probes the host environment (OS, CPU, memory, disk, Python, Node, and which AI agent is present), takes a safety snapshot of key files, checks integrity and auto-repairs missing pieces, installs npm and pip dependencies on demand, generates the .overmind_env.json plus hook and MCP wiring into Claude Code, and runs a smoke test across all six channels with automatic rollback on failure. After install the user sets a DEEPSEEK_API_KEY and restarts Claude Code. The six-channel architecture splits work across CH1 (a 37-stage serial pipeline for intent prediction, safety gates, red-team review, output shielding, and other checks), CH2 (48 parallel broadcast modules), CH3 and CH4 (the z2 hub, where a daemon.py FleetWatcher scans all CC sessions every 5 seconds and broadcasts to peers through .fleet_broadcast.md and an event queue), and CH5 and CH6 (the n2 terminal, where a communicator filters incoming context through a flash model at a claimed 94 percent compression and then post-processes through 8 serial and 11 parallel modules). Memory is layered five ways: semantic memory in SQLite with FTS5 and the jieba Chinese tokenizer, procedural memory that promotes a pattern to a reusable template after three uses, episodic session summaries, a knowledge graph of 4500-plus nodes connected by ten relation types like depends_on and blocked_by, and a feedback loop that scores effectiveness from 4500-plus events. Twenty MCP tools expose search, save, graph expansion, fleet status, and skill management. The README claims a 0.001 ms p50 core latency, 270 pipeline stages per second, under 500 MB memory for all 66 modules, daily API cost of 5 to 15 US cents, and a 42-of-42 release self-test. A competitive table at the end compares it to Mem0 and AgentMemory and pitches Overmind as the brain rather than the hard drive of an AI agent.