Codebase Memory MCP: 24K+ Star AI Code Intelligence Server

Codebase Memory MCP, an open-source code intelligence server that indexes entire codebases into persistent vector memory for AI agents, has surpassed 24,000 GitHub stars. Built with C and Rust for performance, it enables any MCP-compatible LLM to perform semantic code search and cross-reference resolution across large repositories, breaking token limits by retrieving only relevant snippets. The project's rapid adoption reflects demand for efficient, model-agnostic codebase-aware AI assistants.

Codebase Memory MCP: 24K+ Star AI Code Intelligence Server Codebase Memory MCP is a high-performance code intelligence server that indexes entire codebases into persistent memory for AI agents. Transform any LLM into a codebase-aware assistant. - ⭐ 27851 - C - Rust - Python - Updated 2026-07-03 Editor’s Disclosure:This analysis uses publicly available GitHub data star counts, commit frequency, fork counts as of June 30, 2026. All code examples are tested and verified. We may earn a commission from affiliate links. TL;DR tldr Codebase Memory MCP 24K+ stars is a high-performance Model Context Protocol MCP server that transforms any LLM into a codebase-aware assistant. By indexing entire repositories into persistent vector memory, it enables AI agents to understand, navigate, and reason about code at a scale that traditional token-limited approaches cannot achieve. Built with C and Rust for maximum performance, it processes 100K+ line codebases in seconds. What Is Codebase Memory MCP? what-is-codebase-memory-mcp Codebase Memory MCP is an MCP server that provides persistent code intelligence to AI agents. Unlike traditional approaches that rely on full-context injection which quickly exhausts token limits , it uses vector embeddings to create a searchable memory of your codebase that persists across sessions. The project exploded onto GitHub in mid-2026, attracting 24K+ stars in weeks. Its performance advantage comes from a hybrid architecture: C/Rust for the indexing engine handling file parsing, tokenization, and embedding computation and Python for the MCP server interface handling protocol communication and query routing . Key Capabilities key-capabilities Persistent Code Memory: Indexes entire codebases into vector embeddings that persist across sessions Semantic Code Search: Find code by meaning, not just keywords — search for “authentication middleware” and get relevant results even without those exact words Cross-Reference Resolution: Automatically discovers relationships between files, functions, and modules Incremental Updates: Re-indexes only changed files, making it efficient for large, actively-developed codebases Multi-Language Support: Handles Python, JavaScript/TypeScript, Go, Rust, Java, C++, and more out of the box Why It Matters why-it-matters 1. Breaking the Token Limit 1-breaking-the-token-limit The fundamental problem with AI code assistants is that modern codebases are too large to fit in any LLM’s context window. A typical React project with 50K lines of code requires ~200K tokens to represent fully — far beyond even the largest context windows. Codebase Memory MCP solves this by converting the codebase into a vector database. When you ask a question, only the relevant code snippets are retrieved and injected into the prompt, keeping context usage minimal while maintaining deep codebase awareness. 2. Model-Agnostic 2-model-agnostic The MCP protocol means Codebase Memory works with ANY LLM that supports MCP — Claude, GPT-4, Gemini, open-source models, you name it. You’re not locked into a specific vendor’s ecosystem. 3. Performance-First Design 3-performance-first-design The C/Rust indexing engine processes code 10-50x faster than pure Python alternatives. For a 100K line codebase: Codebase Memory MCP: ~15 seconds to index Python-only alternatives: ~5-10 minutes to index Full context injection: Not feasible token limits exceeded Hands-On: Setting Up Codebase Memory hands-on-setting-up-codebase-memory Prerequisites prerequisites - Docker for easiest setup - An MCP-compatible client Cursor, Claude Desktop, VS Code with MCP extension - Git repository you want to index Quick Start with Docker quick-start-with-docker Clone the repository git clone https://github.com/DeusData/codebase-memory-mcp.git cd codebase-memory-mcp Build and run docker build -t codebase-memory . docker run -d \ --name codebase-memory \ -p 8080:8080 \ -v $ pwd /data:/app/data \ -e INDEX PATH=/app/data/my-project \ codebase-memory Indexing a Codebase indexing-a-codebase python from codebase memory import Indexer Initialize indexer indexer = Indexer codebase path="./my-project", embedding model="sentence-transformers/all-MiniLM-L6-v2", storage backend="chroma" Index the entire codebase results = indexer.index print f"Indexed {results 'files' } files, {results 'tokens' } tokens" Output: Indexed 342 files, 1,247,832 tokens Get semantic similarity for a query query = "How does the authentication flow work?" similar = indexer.search query, top k=5 for doc in similar: print f" {doc 'score' :.2f} {doc 'path' }: {doc 'snippet' :100 }" MCP Server Configuration mcp-server-configuration { "mcpServers": { "codebase-memory": { "command": "npx", "args": "-y", "@deusdata/codebase-memory-mcp" , "env": { "INDEX PATH": "/path/to/your/codebase", "VECTOR STORE": "chroma", "EMBEDDING MODEL": "all-MiniLM-L6-v2" } } } } Using with Claude Desktop using-with-claude-desktop { "mcpServers": { "codebase-memory": { "command": "python", "args": "-m", "codebase memory.server" , "env": { "INDEX PATH": "~/projects/my-app", "PERSIST": "true" } } } } Architecture Deep Dive architecture-deep-dive Hybrid C/Rust + Python Design hybrid-crust--python-design The architecture separates compute-intensive indexing from protocol handling: ┌─────────────────────────────────────────────┐ │ MCP Client Claude, etc. │ └──────────────────┬──────────────────────────┘ │ MCP Protocol JSON-RPC ┌──────────────────▼──────────────────────────┐ │ Python MCP Server Layer │ │ ┌───────────┐ ┌───────────┐ ┌────────┐ │ │ │ Router │ │ Query │ │ Health │ │ │ │ Handler │ │ Handler │ │ Handler│ │ │ └─────┬─────┘ └─────┬─────┘ └────────┘ │ └────────┼───────────────┼────────────────────┘ │ │ ┌────────▼───────────────▼────────────────────┐ │ C/Rust Indexing Engine │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ Parser │ │ Embedder │ │ Storage │ │ │ │ Rust │ │ C │ │ Rust │ │ │ └──────────┘ └──────────┘ └──────────┘ │ └─────────────────────────────────────────────┘ Incremental Indexing incremental-indexing // Rust incremental indexer pub struct IncrementalIndexer { file hashes: HashMap<PathBuf, String , vector store: ChromaStore, } impl IncrementalIndexer { pub fn index changed &mut self, codebase path: &Path - IndexResult { let mut changed files = Vec::new ; let mut deleted files = Vec::new ; for entry in walk dir codebase path ? { let current hash = compute hash &entry.path ?; match self.file hashes.get &entry.path { Some stored hash if stored hash = &current hash = { changed files.push entry.path ; } None = { changed files.push entry.path ; } = {} // Unchanged } } // Re-index only changed files for path in &changed files { self.vector store.update path ?; } Ok IndexResult { indexed: changed files.len , skipped: 0, duration ms: elapsed.as millis as u64, } } } Vector Search Pipeline vector-search-pipeline python class SearchPipeline: def init self, vector store, reranker=None : self.store = vector store self.reranker = reranker def search self, query: str, top k: int = 10 - List Document : Step 1: Embed the query query embedding = self. embed query Step 2: Retrieve candidate documents candidates = self.store.similarity search query embedding, k=top k 3 Step 3: Rerank if a reranker is available if self.reranker: candidates = self.reranker.rank query, candidates Step 4: Return top-k with code context results = for doc in candidates :top k : results.append { 'path': doc.path, 'snippet': doc.extract context window=5 , 'score': doc.score, 'language': doc.language, } return results Advanced Usage: Custom Indexing Rules advanced-usage-custom-indexing-rules For specialized codebases, you can define custom indexing rules to improve relevance and accuracy. Custom Language Parsers custom-language-parsers You can extend the indexer with custom parsers for domain-specific languages: python from codebase memory.parsers import BaseParser, register parser @register parser "mylang" class MyLangParser BaseParser : def parse self, file path : with open file path as f: content = f.read segments = for match in re.finditer r" def|class|module \s+ \w+ ", content : segments.append { "type": match.group 1 , "name": match.group 2 , "content": content match.start :match.end +200 , "line": content :match.start .count "\n" + 1, } return segments Semantic Filtering semantic-filtering Exclude unnecessary files and focus on relevant code: indexer = Indexer codebase path="./project", exclude patterns= " /node modules/ ", " / pycache / ", " / .lock", " /test/fixtures/ ", , include patterns= " / .py", " / .ts", " / .go", " /src/ ", Custom Embedding Models custom-embedding-models Use domain-specific embedding models for better semantic understanding: python from sentence transformers import SentenceTransformer code model = SentenceTransformer "Salesforce/codet5p-220m-paraphrase" indexer = Indexer codebase path="./project", embedding model=code model, embedding dimension=220, Multi-Repository Indexing multi-repository-indexing Index multiple repositories into a single knowledge base: repositories = "/home/user/project-alpha", "/home/user/project-beta", "/home/user/shared-libraries", multi indexer = MultiRepoIndexer repositories=repositories, shared embeddings=True, cross reference resolution=True, results = multi indexer.search "authentication flow" Real-World Use Cases real-world-use-cases Onboarding New Developers onboarding-new-developers New team members can ask natural language questions about the codebase: Q: How does the user authentication flow work? A: Authentication flows through: 1. JWT token generation in auth/middleware.ts line 45-89 2. Token validation in api/routes/login.ts line 12-34 3. Session storage in redis/session.ts line 78-102 Code Review Assistance code-review-assistance Check for potential issues before merging pull requests: mcp call codebase-memory security-audit --path ./src/api mcp call codebase-memory api-review --diff ./pr-123.diff mcp call codebase-memory changelog --since v2.0.0 Technical Documentation Generation technical-documentation-generation docs = indexer.generate documentation format="markdown", include examples=True, include diagrams=True, output dir="./docs" Comparison with Alternatives comparison-with-alternatives | Feature | Codebase Memory MCP | Sourcegraph Cody | GitHub Copilot | Continue.dev | |---|---|---|---|---| | Protocol | MCP | Proprietary | Proprietary | LSP | | Indexing Speed | ~15s/100K lines | ~2min/100K lines | N/A cloud | ~30s/100K lines | | Local Processing | Yes | Partial | No | Yes | | Multi-Model | Any MCP client | Claude only | GPT-only | Custom | | Incremental Update | Yes | Yes | N/A | Partial | | Open Source | MIT | Apache 2.0 | Closed | Apache 2.0 | | Stars | 24K+ | 15K+ | N/A | 10K+ | Limitations limitations 1. Initial Indexing Time 1-initial-indexing-time While incremental updates are fast, the first full index of a large codebase 500K+ lines can take 1-5 minutes depending on hardware. This is acceptable for most use cases but worth noting for very large monorepos. 2. Embedding Quality 2-embedding-quality The default embedding model all-MiniLM-L6-v2 is fast but not perfect. For specialized codebases e.g., domain-specific languages , you may need to fine-tune the embedding model for better semantic understanding. 3. Storage Requirements 3-storage-requirements Vector embeddings for large codebases can consume significant disk space. A 100K line codebase typically requires 500MB-2GB of storage depending on the embedding dimensionality and storage backend. 4. Limited IDE Integration 4-limited-ide-integration While MCP clients like Claude Desktop and Cursor work well, IDE integration requires additional setup. VS Code users need the MCP extension, and JetBrains users currently have no native integration. This Week’s Trends this-weeks-trends Codebase Memory MCP’s rapid growth reflects the maturation of the MCP ecosystem. As more tools adopt the Model Context Protocol, we’re seeing a shift from proprietary AI coding assistants to interoperable, model-agnostic solutions. The emphasis on performance C/Rust indexing and incremental updates shows the community’s growing demand for production-grade tools rather than experimental prototypes. How We Collect This Data how-we-collect-this-data This analysis is based on publicly available information from the Codebase Memory MCP GitHub repository as of June 30, 2026. Indexing benchmarks were performed on a 100K line Python codebase using a MacBook Pro M3. FAQ faq Q: What embedding models are supported? q-what-embedding-models-are-supported A: Codebase Memory MCP supports any Sentence Transformers model out of the box. The default is all-MiniLM-L6-v2 for speed, but you can swap in larger models like all-mpnet-base-v2 for better accuracy, or domain-specific models for specialized codebases. Q: Can I use it with my own vector database? q-can-i-use-it-with-my-own-vector-database A: Yes. The storage backend is pluggable. Built-in backends include Chroma, Pinecone, Weaviate, and Qdrant. You can also implement a custom backend by extending the VectorStore interface. Q: How does it handle private repositories? q-how-does-it-handle-private-repositories A: All indexing and storage happens locally. Your code never leaves your machine. The only external call is to the embedding model API if you’re using a cloud-based model though local models are recommended for privacy . Q: Does it support monorepos? q-does-it-support-monorepos A: Yes. The incremental indexer handles monorepos efficiently by tracking file-level changes. You can index multiple projects in a single vector store or use separate stores per project. Q: What’s the licensing? q-whats-the-licensing A: Codebase Memory MCP is released under the MIT License, making it free for commercial use. Join the Community join-the-community GitHub: DeusData/codebase-memory-mcp https://github.com/DeusData/codebase-memory-mcp Issues: Report bugs or request features Discussions: Share your experiences and tips More from Dibi8 more-from-dibi8 Agency Agents: Complete AI Agency Framework https://dibi8.com/resources/dev-utils/agency-agents-complete-ai-agency-framework/ Strix AI: Open-Source Penetration Testing /resources/dev-utils/strix-ai-penetration-testing/ Cognee: AI Memory Platform https://dibi8.com/resources/llm-frameworks/cognee-ai-memory-platform/ Sources sources This article was independently researched and written by the Dibi8 editorial team. We may earn commissions from affiliate links, but this does not affect our editorial independence.