memento-ai

Your codebase, remembered.

memento watches your git commits and maintains a living memory of your project — readable by both humans and AI.

How it works

git commit → memento analyzes the diff → updates markdown memory files

Memory files live in .memento/memory/ — plain markdown, diffable, committable to your repo.

Quick start

pip install memento-ai
cd your-project
memento init
memento process --all

That's it. Your project now has memory.

Ask questions

memento ask "what does the auth system do?"
memento ask "what changed in the last few commits?"

Features

• Incremental — only processes new commits
• Offline-first — works with Ollama or embedded local models (zero API cost)
• Multi-provider — OpenAI, Anthropic, Claude CLI, Ollama, local inference
• Human-readable — memory is plain markdown
• Git-native — auto-processes via post-commit hook
• Private — your code never leaves your machine (with local providers)

LLM Providers

Using Ollama (recommended for privacy)

ollama pull qwen2.5-coder:7b

Edit .memento/config.toml:

[llm]
provider = "openai"
model = "qwen2.5-coder:7b"
base_url = "http://localhost:11434/v1"

Using embedded local model (zero setup)

pip install memento-ai[local]

Edit .memento/config.toml:

[llm]
provider = "local"
# Auto-downloads Qwen2.5-Coder-3B (~2GB) on first run

Using Claude CLI (free with Claude Max/Pro)

[llm]
provider = "claude-cli"

Requires Claude Code installed and logged in.

Commands

Configuration

.memento/config.toml:

[llm]
provider = "claude-cli"       # openai, anthropic, claude-cli, local
model = "gpt-4o-mini"         # model name (provider-specific)
base_url = "https://..."      # API base URL (openai provider)
temperature = 0.3
max_tokens = 2048

[processing]
chunk_size = 4000             # max diff size before chunking
summary_every = 10            # regenerate summary every N commits
ignore_patterns = ["*.lock", "dist/*"]

[memory]
dir = "memory"                # subdirectory for memory files
max_module_size = 5000        # max lines per module

Memory structure

.memento/
├── config.toml          # your configuration
├── state.json           # processing state (gitignored)
└── memory/
    ├── SUMMARY.md       # auto-generated project overview
    ├── api-endpoints.md # module: API routes and patterns
    ├── auth-system.md   # module: authentication logic
    └── database.md      # module: schema and queries

Modules are created and maintained automatically based on what the LLM finds in your commits.

MCP Server

memento exposes project memory via the Model Context Protocol, so any MCP-compatible AI tool can access your project memory automatically.

pip install "memento-ai[mcp]"

Claude Code

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "memento": {
      "command": "memento-mcp"
    }
  }
}

Cursor

Add via Settings → MCP Servers:

{
  "mcpServers": {
    "memento": {
      "command": "memento-mcp"
    }
  }
}

Available MCP tools

Available MCP resources

Export

Export project memory to files that AI tools read automatically:

memento export --format claude   # → CLAUDE.md
memento export --format cursor   # → .cursor/rules/memento.mdc
memento export --format copilot  # → .github/copilot-instructions.md

License

MIT