AIDOCS: Orchestration Layer and Persistent Memory System for AI Coding Agents

This article introduces AIDOCS—a system that provides a unified orchestration layer for AI coding tools like Claude Code, Codex, OpenCode, and Cursor. It addresses core pain points such as agent context loss and repeated codebase exploration, enabling cross-session persistent memory and multi-agent collaboration. It does not alter existing tools; instead, it builds capabilities on top of them.

With the popularity of AI coding assistants like Claude Code, GitHub Copilot, and Cursor, developers face core pain points: every time a new session starts, the agent needs to "re-familiarize" itself with the entire codebase, leading to token waste, time consumption, and fragmented development experience. AIDOCS aims to solve this problem by acting as an orchestration layer that provides persistent memory and intelligent orchestration capabilities for existing tools.

AIDOCS (AI Documentation and Orchestration System) is an open-source Model Context Protocol (MCP)-compatible runtime developed by the CodeNexus team and released under the Apache 2.0 license. It is positioned as an enhancement tool for existing agents, not a replacement, integrating mainstream coding tools to provide infrastructure such as memory persistence and task orchestration. The core architecture consists of three layers: Core Layer (portable Markdown memory, routing rules, etc.), MCP Runtime Layer (indexing, retrieval, orchestration, host integration), and Dashboard Layer (Tauri desktop application for graphical monitoring and control).

AIDOCS's core mechanisms include: 1. Routing-based memory initialization: Gradually loads required context via a startup chain to avoid blindly dumping the entire repository; memory files are stored in the /.MEMORY/ path using pure Markdown format (e.g., INDEX.md, session files, etc.), which is open, transparent, and version control-friendly. 2. Conductor Orchestrator: Long-living, responsible for task distribution, Lane Agent management, conflict resolution, supports session isolation, and routes to appropriate models based on task type during multi-agent collaboration (e.g., Claude for code generation, GPT for refactoring, etc.). 3. Index-based retrieval: Advanced retrieval based on symbols, semantics, and relationships, supporting multi-module/project indexing for large monorepos.

AIDOCS has built-in complete security gating mechanisms: heuristic referee (assesses operation risk levels), output guard (blocks sensitive information leakage), tool policies (fine-grained control over available tool sets), and RBAC permission model (role-based access control). These security layers are deeply integrated with orchestration logic to ensure multi-agent collaboration behaviors are predictable, auditable, and rollbackable.

AIDOCS adopts a "zero migration" concept, working as an overlay layer on top of existing tools. Installation methods are diverse: Python package (pip install aidocs-mcp && aidocs setup), Windows installer (download AIDOCS-Setup.exe from Releases), macOS/Linux one-click script (curl command). After configuration, enter /aidocs in the IDE to activate managed mode. Common commands include /aidocs (start/resume session), /reingest (refresh memory), /archive (archive work to CHANGELOG), and /clean (clean up code).

AIDOCS supports a multi-language classification system. User intent descriptors for different languages are defined via TOML files in the action_tokens/ directory. Currently, it supports English, Italian, etc., and specific languages can be enabled in aidocs.toml (e.g., [languages] enabled="en"). System behavior can be customized via configuration files: aidocs.toml (static definitions like log limits, directories to skip for indexing), action_hooks/*.toml (interactive text templates), and aidocs-plugin.json (OpenCode plugin behavior).

AIDOCS represents the evolution direction of AI coding tools: shifting from "single-session intelligence" to "persistent organizational intelligence", solving the context continuity problem while being compatible with existing toolchains. For individual developers: a more coherent coding experience, where agents remember project structures and preferences; for teams: shared collective memory, enabling new members to quickly understand project conventions; for enterprises: security gating and auditing meet compliance requirements. As the MCP ecosystem matures, orchestration layers like AIDOCS will become standard components. Its open-source strategy (Apache 2.0) and active testing (1287 use cases) lay the foundation for community contributions and long-term maintenance.