GUAN Framework: Innovative Practice of Multi-LLM Collaboration and Persistent Cognitive Management

The GUAN Framework is an open-source framework built on cognitive science theories, designed to address context forgetting and subscription waste in AI-assisted development. It enables cross-model persistent cognitive configuration via native Git mechanisms and supports multi-LLM collaboration.

With the improvement of large language model capabilities, developers face two prominent issues when using AI tools:

1. Context Forgetting: Each new session starts from scratch, requiring repeated explanations of project background and tried solutions, which wastes time and easily leads to information omission;
2. Subscription Waste: Developers subscribe to multiple AI services but leave them idle due to high context switching costs, resulting in unreasonable resource allocation.

The GUAN Framework is built on three cognitive science theories:

1. Extended Mind Theory: Cognitive processes extend to the external environment; developers' cognitive configuration files are regarded as mind extensions, allowing AI models to load instantly for seamless context continuation;
2. Scaffolding and Replacement: AI should enhance human thinking rather than replace it; the Challenge Contract Protocol ensures AI assistants do not weaken developers' independent judgment;
3. Hollowing Warning: Built-in challenge mechanisms prevent AI from eroding users' independent thinking ability and avoid cognitive hollowing.

The GUAN Framework adopts a file-based context system, with configurations stored in Git repositories for version control and collaboration:

• Parallel Session Protocol: The session_id + slot mechanism solves multi-window conflicts, with each window's unique ID embedded in the file name;
• Challenge Contract Protocol v1.2: Extended to 8 trigger conditions (e.g., batch overload, requirement conflicts, etc.), which automatically trigger reviews;
• Semi-automatic Cognitive Collection: AI monitors cognitive value signals and prompts users to save as cognitive cards;
• Quality Filter: Candidate insights must meet 4 conditions to exclude temporary data, ensuring the value density of cognitive configurations;
• Multi-LLM Orchestration: Claude (70-80%, Commander/Executor), Codex (15-20%, Reviewer/Builder), Gemini (5-10%, Researcher/Analyst) collaborate, with external agents automatically called via Trigger Matrix v1.2.

The GUAN Framework originated from the real scenario of an independent developer managing complex enterprise systems (using up the Claude Max quota in 3 days while other subscriptions were idle). Four constraints are considered in its design:

1. Cognitive Load Management: File-based configuration reduces memory burden;
2. Cost Control: Intelligently assign tasks to models at different price points;
3. Quality Assurance: Multi-layer reviews prevent error accumulation;
4. Security Boundaries: 9 absolute prohibition rules ensure the safety of multi-LLM orchestration.

The framework design references several cutting-edge studies:

• Stanford Digital Twin Study (2024): 2-hour interviews can achieve 85% behavioral accuracy, verifying the feasibility of the guided approach;
• Stanford SCALE Study (2025): Digital twin responses are more predictable; explanations require the Challenge Contract Protocol to introduce questioning;
• Columbia Business School Study (2025): Detailed character descriptions amplify AI bias, so GUAN cards use atomic statements instead of narrative descriptions.

The GUAN Framework represents a new paradigm for AI collaboration: shifting from single-model dependence to multi-model synergy, and from conversational interaction to persistent cognitive management. In the future, it will become a standard component of the developer toolchain—not just a technical tool, but a work philosophy that extends human thinking capabilities. Its open-source nature allows the community to continuously improve it to adapt to technological changes.