KimGLee/Cambium

What it solves

Cambium provides a governance standard and a set of reference tools to manage knowledge repositories that are maintained by LLM agents. It prevents agents from "guessing" during interrupted tasks, ensures that work is performed according to specific rules, and clearly separates the decisions that belong to a human operator from those made by the agent.

How it works

Cambium operates through a layered architecture consisting of a normative kernel (governance semantics), selected profiles (repository-specific scope and rules), and runtime state (the current progress of work). It uses a system of "Cards" (checklists) and "Read Sets" (declarations of required material) to guide agents.

To maintain a strict audit trail, it employs three runtime ledgers—Coverage, Required Queue, and Progress—to track which objects exist, what work is pending, and the current task contract. Changes to shared state are handled by controlled writers to ensure revisions and receipts are recorded deterministically.

Who it’s for

It is designed for operators who use LLM agents to maintain large-scale knowledge bases or repositories, requiring strict governance, resumability of long-running tasks, and a deterministic audit trail of agent actions.

Highlights

• Resumable Work: Uses persistent state and ledgers to allow interrupted tasks to resume without loss of context or guesswork.
• Deterministic Tools: Provides a suite of stable public commands for controlled writes, schema checks, and state transitions.
• Host-Neutral Interface: Generates configurations for various agent hosts like Claude Code, Codex, and Kimi Code via an MCP projection.
• Strict Governance: Separates the normative kernel rules from adopter-owned profiles, ensuring that repository-specific policies are enforced.
• Append-only Receipts: Maintains a history of actions through JSONL receipts and Terminal Proof bindings for verification.