CodeGraph Memory MCP Lite

CodeGraph Memory MCP Lite

Local-first MCP-style code memory for research repositories: file search, Python symbols, import graph, and repo context packs.

  

<p align="center"> <img src="assets/screenshots/mcp-reference.png" alt="Open-source developer tooling public page screenshot" width="880"> </p>

_{Image source: public Model Context Protocol homepage screenshot, https://modelcontextprotocol.io/. Used as a visual reference for local code-indexing and agent-tool workflows.}

Research codebases are often small enough that a heavyweight code intelligence stack is overkill, but large enough that agents lose track of modules, configs, training scripts, and README claims. This project provides a tiny local codegraph memory layer for AI agents. It indexes a repository, extracts Python classes/functions/imports with the standard ast module, stores file text in SQLite FTS, and exposes MCP-style tools over stdio.

If this helps your agent understand a research repo faster, a star helps other builders discover it.

Features

• Local-only SQLite database.
• File full-text search over code, Markdown, configs, and scripts.
• Python symbol extraction: classes, functions, async functions, and imports.
• Lightweight import graph and symbol lookup.
• MCP-style tools:
• index_codebase
• search_code_memory
• find_symbol
• get_dependency_graph
• summarize_repo_context

Quick Start

git clone https://github.com/StaryMoon/codegraph-memory-mcp-lite.git
cd codegraph-memory-mcp-lite
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

codegraph-memory index examples/sample_repo
codegraph-memory symbol Trainer
codegraph-memory graph --limit 20
codegraph-memory context

MCP Config

{
  "mcpServers": {
    "codegraph-memory-lite": {
      "command": "python3",
      "args": ["-m", "codegraph_memory_mcp_lite.server", "serve"],
      "env": {
        "CODEGRAPH_MEMORY_DB": "/absolute/path/to/codegraph-memory.sqlite"
      }
    }
  }
}

Why This Exists

Most code assistants can read files, but they often lack a persistent local map of:

• where classes and functions are defined;
• which modules import which modules;
• what README claims the repo makes;
• what training/evaluation scripts exist;
• which files are relevant to a user query.

This project gives an agent a small memory substrate before it starts editing.

Related Projects

• ai-researcher-skills
• paper-memory-mcp-lite
• obsidian-research-brief-kit

License

MIT.