TokenRollAI/llmdoc

TokenRoll LLMDoc for Coding Agent

What it solves

Coding agents often struggle to understand a repository's architecture, design decisions, and cross-module contracts because this high-level context is rarely fully captured in the source code itself. This leads to agents having to rediscover the architecture every session, which is inefficient and expensive in terms of token usage.

How it works

llmdoc creates a persistent, durable knowledge map in a dedicated llmdoc/ directory within the repository. Instead of duplicating the source code, it stores "engineering meaning"—decisions, constraints, and invariants.

It operates through two layers:

  1. Agent Workflows: High-level roles (investigator, reflector, recorder) that handle the judgment and maintenance of the knowledge base.
  2. Runtime CLI: A deterministic tool for retrieval and mechanics, allowing agents to search, index, and retrieve specific context based on the task at hand.

When a code change occurs, the system identifies affected knowledge and creates a review obligation for the agent to verify the facts rather than automatically rewriting documentation.

Who it’s for

Developers using AI coding agents (such as Claude Code or Codex) who want to maintain a long-term, structured architectural context that persists across sessions.

Highlights

  • Task-scoped retrieval: Retrieves only the smallest useful set of knowledge needed for a specific task to save tokens.
  • Semantic maintenance: Verifies knowledge as code evolves, preventing the accumulation of stale notes.
  • Plugin integration: Native plugins for Claude Code and Codex, with a portable integration recipe for other agents.
  • Knowledge boundaries: Strictly separates stable knowledge (llmdoc/) from temporary investigations (.llmdoc-tmp/).

Related

  • Project
    JamesShi96/project-butler

    A persistent memory system for AI coding assistants that uses a structured Markdown-based memory stack to maintain project context, rules, and progress across sessions.

  • Project
    mindmuxai/brain.md

    A persistent, file-based memory layer for coding agents that stores project decisions and constraints as plain Markdown in the repository.

  • Project
    Autoloops/greplica

    A persistent, searchable engineering memory system for AI coding agents that reduces redundant context-gathering and token usage by storing durable repository knowledge.

  • Project
    rohitg00/agentmemory

    A persistent memory layer for AI coding agents that captures and compresses session data to eliminate the need for re-explaining project context across sessions and tools.