mex-memory/mex

Persistent project memory for AI coding agents. Structured scaffold + drift detection CLI.

What it solves

AI coding agents often struggle with large codebases because they either rediscover the same architectural patterns every session or flood their context windows with massive, stale instruction files. mex solves this by creating a "living wiki"—a repo-local, structured Markdown documentation system that grows and evolves alongside the code, ensuring agents have precise, task-relevant context without redundant repository scans.

How it works

mex uses a combination of a deterministic local code graph and structured Markdown files to maintain project knowledge:

  1. Code Mapping: It builds a local graph using Tree-sitter and SQLite to index symbols and relationships across languages like TypeScript, Python, and Rust.
  2. Wiki Generation: Agents use this graph to populate a structured Markdown wiki (e.g., architecture, conventions, and patterns) stored directly in the repository.
  3. Context Routing: Instead of loading everything, agents use a routing system (ROUTER.md) to load only the specific wiki pages and code-graph neighborhoods relevant to the current task.
  4. Drift Detection: The tool identifies when code changes invalidate wiki claims through "grounding" (linking claims to exact code symbols), allowing agents to perform targeted repairs via mex sync.

Who it’s for

Developers using AI coding agents (such as Claude Code, Cursor, Windsurf, or GitHub Copilot) who want to reduce token usage, lower latency, and prevent their AI assistants from losing architectural context between sessions.

Highlights

  • Grounded Knowledge: Connects documentation claims to exact code symbols, allowing the system to detect when refactors break the documentation.
  • Task-Aware Routing: Reduces context window bloat by loading only the necessary documentation and source code for a specific job.
  • Deterministic Retrieval: Uses a code graph to provide source-backed retrieval and impact analysis without spending AI tokens on mechanical search.
  • Multi-Agent Compatibility: Provides project anchors for various tools like .cursorrules or CLAUDE.md to integrate seamlessly with different AI coding assistants.

Related

  • Project
    AlmanacCode/codealmanac

    A living wiki for codebases that uses AI agents to automatically extract and maintain architectural context and system knowledge in Markdown files within the repository.

  • Project
    agent0ai/dox

    A lightweight framework using a hierarchy of AGENTS.md files to give AI agents precise project context and prevent blind edits during coding.

  • Project
    akitaonrails/ai-memory

    A long-term memory system for AI coding agents that creates a shared, persistent Markdown wiki from session observations to enable seamless context handoffs between different agents.

  • Project
    langchain-ai/openwiki

    A CLI tool that uses AI agents to create and maintain a self-updating Markdown wiki for codebases or personal knowledge, featuring grounded claims to ensure factual accuracy.

  • Project
    vercel-labs/lat.md

    A knowledge graph system for codebases that uses interconnected markdown files and source code annotations to provide AI agents and humans with scalable, searchable context.