Undefined

Undefined is a high-performance async bot platform for QQ group and private chat scenarios, differentiated by its cognitive memory architecture. The memory system has three tiers: short-term memory maintains conversation continuity, cognitive memory enables AI to proactively observe and extract facts then asynchronously rewrite and store them in a vector database (with semantic search, time-decay weighting, MMR deduplication, cross-group linkage, and auto-profiling), and pinned memos provide AI self-constraints and todo management.

The platform includes 6 built-in agents (info, web search, file analysis, code analysis, entertainment, code delivery) covering 64+ tool capabilities, with fine-grained inter-agent tool sharing and whitelist control via callable.json. The skills system supports hot-reload, and 11 toolset categories are extensible on demand.

The architecture employs an 8-layer design with a "station-train" queue model, where 4-priority scheduling ensures high-value requests are served first. It supports multi-model pool concurrent comparison, MCP protocol for external tool integration, Anthropic Skills standard for domain knowledge injection, local knowledge base, global meme library, and cron-based scheduled tasks. The management plane provides a Management-first WebUI and cross-platform console with separated Management API and Runtime API, allowing full initialization even when configuration is incomplete.

Key Capabilities

• Three-tier memory: short-term / cognitive (vector search + time-decay + cross-group linkage) / pinned memos
• 6 built-in agents: info_agent, web_agent, file_analysis_agent, naga_code_analysis_agent, entertainment_agent, code_delivery_agent
• 11 toolset categories: group, messages, memory, contacts, group_analysis, notices, render, scheduler, cognitive, mcp, memes
• Multi-model pool: round-robin / random / user-specified, with concurrent comparison
• Four independent model configs: chat / vision / agent / security
• Local knowledge base: text vectorized into ChromaDB with keyword/semantic search
• MCP protocol: agents can have private MCP configs for external tools
• Scheduled tasks: Crontab syntax with self_instruction mode
• Access control: group/private chat blacklist/allowlist
• Security: built-in independent security model for injection and malicious content detection

Installation

Prerequisites: Python >= 3.11, < 3.14; uv (recommended) or pip; OneBot V11 endpoint (e.g. NapCat / Lagrange.Core); AI model API (OpenAI-compatible).

git clone --recursive https://github.com/69gg/Undefined.git
cd Undefined
pip install uv
uv sync
uv run playwright install
uv run Undefined-webui

Entry Commands

• uv run Undefined — start bot directly
• uv run Undefined-webui — start management console (recommended)

Core Config (config.toml)

• Basic: core.bot_qq, core.superadmin_qq, onebot.ws_url
• Models: models.chat / models.vision / models.agent / models.security
• Cognitive memory: cognitive.enabled, cognitive.query.*, models.embedding.*
• WebUI: webui.url, webui.port, webui.password
• MCP: mcp.config_path