Agent Context

Agent Context Architecture

[](LICENSE) [](https://github.com/lx-wnk/Agent-Context/releases/latest) [](https://github.com/lx-wnk/Agent-Context/actions/workflows/ci.yml)

A project-based setup and memory-handling system for AI coding agents, with first-class Claude Code support. Its entry point is the agent-agnostic `AGENTS.md`. Optimized for structuring project knowledge so that your agent always has the right context at the right time — without bloating the context window.

Instead of dumping everything into a single `CLAUDE.md`, Agent Context provides a layered architecture: all layers (0-3) are loaded at startup via `@`-includes in `AGENTS.md`, keeping the baseline at ~150-200 lines. Detailed reference (skills, memory files) is pulled in on-demand based on the task at hand. Auto-updates keep shared infrastructure current across all your projects.

Contents

• The Problem
• The Solution
• Scope
• Installation
• Architecture
• Documentation
• Contributing
• License

The Problem

Claude Code loads project instructions into its context window every conversation. Most projects dump everything into a single `CLAUDE.md`, resulting in:

• Context bloat: 500-1000+ lines loaded for every task, even a one-line CSS fix
• Duplication: Same information in CLAUDE.md, README.md, .claude/rules/, and memory files
• Noise: Entity schemas, route tables, and file trees that Claude can discover by reading the code
• No structure: Flat files with no way to load context progressively based on the task

The Solution

A layered architecture where all layers load at startup via `@`-includes in `AGENTS.md`