mstuart/code-memory: Persistent memory for AI coding - semantic search, git history, and intelligent context preservation · GitHub

Persistent memory for AI coding - semantic search, git history, and intelligent context preservation

Code Memory is an MCP (Model Context Protocol) server that gives AI coding assistants long-term memory of your codebase through semantic search, git history analysis, and intelligent context preservation.

AI coding assistants forget:

• 🤔 Context between sessions - "Why did we use JWT instead of sessions?"
• 🔍 Architectural decisions - "What was the rationale for this design?"
• 📚 Historical discussions - "We already tried that approach last month"
• 🕸️ Code relationships - "What else depends on this module?"

Code Memory provides:

• 🧠 Semantic search - Find code by meaning, not just keywords
• 📚 Git history analysis - Extract decisions from commit messages and PRs
• 🕸️ Dependency graphs - Understand code relationships across 16 languages
• 📝 Session learning - Learn patterns from your coding sessions
• 💾 Persistent knowledge - Remember important facts across sessions

npm install -g code-memory

# Initialize in your project
cd your-project
code-memory init

# Index your codebase
code-memory reindex

# Start the MCP server
code-memory serve

# Search from CLI
code-memory search "authentication flow"

From the project where you want Code Memory available, choose a scope:

# Share with this project via .mcp.json
claude mcp add --transport stdio --scope project code-memory -- code-memory serve

# Or make it available to all projects via ~/.claude.json
claude mcp add --transport stdio --scope user code-memory -- code-memory serve

Verify the connection with claude mcp get code-memory.

Find code by meaning, not just keywords:

code-memory search "user authentication logic"
# Finds auth code even if it doesn't contain the word "user"

How it works:

• Uses fastembed for local embeddings (all-MiniLM-L6-v2)
• No API calls, completely offline
• Understands intent, not just text matching

Fast keyword-based search powered by tantivy:

code-memory search --fulltext "async function"

Extract architectural decisions from your git history:

code-memory trace-decision "why microservices"

Finds:

• Explicit decisions in commit messages
• Architectural choices from PR descriptions
• Rationale and "why" statements
• Refactoring decisions
• Breaking changes

Understand what depends on what:

code-memory find-related "UserService"

Analyzes:

• Import relationships
• Co-change patterns (files changed together)
• Dependency chains
• Most coupled modules

Supported languages: Rust, TypeScript, JavaScript, Python, Go, Java, C++, C#, Ruby, PHP, Swift, Kotlin, Scala, Haskell, Elixir, Clojure

Learn from your coding sessions:

code-memory sessions

Extracts:

• Architectural decisions made in sessions
• Error-and-fix patterns
• Refactoring approaches
• Testing strategies

Remember important facts:

# Via MCP server
remember("We use JWT for auth because of scalability requirements")

Code Memory provides 7 MCP tools for AI assistants:

Search codebase with semantic + full-text search:

search_code({
  query: "authentication flow",
  limit: 10
})

Get detailed explanation of a symbol:

explain_code({
  symbol: "UserService",
  context_lines: 20
})

Find why a decision was made:

trace_decision({
  topic: "why microservices",
  max_results: 5
})

Find related code and dependencies:

find_related({
  symbol: "AuthController",
  relationship_type: "both" // "depends_on" | "depended_by" | "both"
})

Store persistent knowledge:

remember({
  key: "auth-strategy",
  value: "We use JWT for stateless auth across microservices"
})

Manually trigger reindexing:

index_project({
  force: true
})

Retrieve learned patterns:

get_session_patterns({
  pattern_type: "architecture", // or "errors", "refactoring", "testing"
  min_confidence: 0.7
})

Initialize Code Memory in a project:

code-memory init

Creates .code-memory/ directory with:

• config.toml - Configuration
• index/ - Search index
• knowledge.json - Persistent knowledge

Start the MCP server:

code-memory serve

Rebuild the code index:

code-memory reindex           # Incremental
code-memory reindex --force   # Full rebuild

Search from the command line:

code-memory search "query"
code-memory search "query" --lang rust
code-memory search "query" --limit 20

Show index statistics:

code-memory stats

View session patterns:

code-memory sessions
code-memory sessions -n 10            # Top 10
code-memory sessions -c 0.8           # Min confidence 0.8
code-memory sessions -f json          # JSON output

Backup and restore knowledge:

code-memory export knowledge.json
code-memory import knowledge.json

Configuration is stored in .code-memory/config.toml:

[indexing]
# File patterns to index
include = ["**/*.rs", "**/*.ts", "**/*.js", "**/*.py"]

# File patterns to ignore
exclude = ["**/node_modules/**", "**/target/**", "**/.git/**"]

# Maximum file size (in bytes)
max_file_size = 1048576  # 1MB

[search]
# Number of results to return by default
default_limit = 10

# Minimum relevance score (0.0-1.0)
min_score = 0.5

[git]
# Analyze git history for decisions
analyze_history = true

# How far back to look (in days)
history_depth = 365

[embedding]
# Embedding model (local, no API calls)
model = "all-MiniLM-L6-v2"

# Embedding dimension
dimension = 384

code-memory/
├── src/
│   ├── indexer/         # Code indexing with tantivy
│   │   ├── walker.rs    # File system traversal
│   │   ├── parser.rs    # Symbol extraction
│   │   └── code_index.rs
│   ├── search/          # Search engines
│   │   ├── fulltext.rs  # Tantivy full-text search
│   │   ├── semantic.rs  # Fastembed semantic search
│   │   └── hybrid.rs    # Combined ranking
│   ├── git/             # Git history analysis
│   │   ├── history.rs   # Commit parsing
│   │   └── decisions.rs # Decision extraction
│   ├── graph/           # Dependency graphs
│   │   ├── imports.rs   # Import parsing
│   │   └── analyzer.rs  # Graph analysis
│   ├── sessions/        # Session tracking
│   │   ├── tracker.rs   # Event extraction
│   │   └── patterns.rs  # Pattern learning
│   ├── mcp/             # MCP server
│   │   ├── server.rs    # JSON-RPC server
│   │   ├── tools.rs     # Tool handlers
│   │   └── protocol.rs  # MCP protocol
│   └── cli.rs           # CLI interface
└── .code-memory/
    ├── config.toml      # Configuration
    ├── index/           # Search index
    └── knowledge.json   # Persistent facts

• Indexing speed: ~10,000 files/minute
• Search latency: <100ms (full-text), <200ms (semantic)
• Memory usage: ~100MB for 50k files
• Binary size: 8MB (including embedding model)

• Free: Up to 5,000 files, basic search
• Pro ($20/month):
  • Unlimited files
  • Advanced query optimization
  • Team knowledge sharing
  • Priority support

Full support (symbol extraction + imports):

• Rust, TypeScript, JavaScript, Python, Go
• Java, C++, C#, Ruby, PHP
• Swift, Kotlin, Scala, Haskell, Elixir, Clojure

Additional languages supported for full-text search only.

Code Memory understands meaning, not just text:

• "user auth" finds authentication code even without those exact words
• Extracts architectural decisions from git history
• Understands code relationships across files
• Learns from your coding sessions

No. Everything runs locally:

• Embeddings generated on your machine (fastembed)
• Search index stored locally (tantivy)
• No network calls during normal operation
• Your code never leaves your computer

Approximately:

• ~50MB per 10,000 files indexed
• Embedding model: 90MB (downloaded once)
• Total: ~150-300MB for a typical project

Yes! Code Memory is an MCP server, so it works with any MCP-compatible tool:

• Claude Code CLI
• Any editor with MCP support
• Custom integrations via MCP protocol

Code Memory only accesses files you explicitly index. It never:

• Uploads code to remote servers
• Shares data with third parties
• Requires authentication or accounts (for free tier)

# Check for errors
code-memory reindex --verbose

# Force rebuild
code-memory reindex --force

# Check config
cat .code-memory/config.toml

# Verify index exists
code-memory stats

# Rebuild index
code-memory reindex --force

# Check file patterns in config

# Check if port is in use
lsof -i :8080

# Start with verbose logging
code-memory serve --verbose

git clone https://github.com/mstuart/code-memory.git
cd code-memory
cargo build --release

cargo test                    # Run all tests
cargo test --lib             # Unit tests only
cargo test --test mcp_tools  # Integration tests

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new features
4. Ensure all tests pass
5. Submit a pull request

MIT License - see LICENSE for details

• 📖 Documentation
• 🐛 Issues
• 💬 Discussions

Built with:

• tantivy - Full-text search
• fastembed - Local embeddings
• git2 - Git integration
• petgraph - Dependency graphs
• tree-sitter - Code parsing

Give your AI assistant a memory. Never lose context again. 🧠

npm install -g code-memory