amygdala 0.2.0

AI coding assistant memory system — git-integrated file summaries

Amygdala

AI coding assistants start every session with amnesia. Amygdala is a git-integrated memory system that tracks file summaries at configurable granularity levels, detects dirty files via git diff, and injects relevant context at session start.

• Provider-agnostic -- Anthropic, OpenAI, Google Gemini, and Ollama supported out of the box
• Adapter system -- pluggable integration with Claude Code, Cursor, Windsurf (Claude Code ships first)
• Git-native -- summaries tracked alongside your code, dirty detection via content hashing
• Extension profiles -- opt-in to framework-specific file types (Unity, Unreal, Python, Node, React, Next.js)
• Auto-capture -- MCP-driven summary refresh using your Claude Code subscription (no API key needed)

Installation

From PyPI

pip install amygdala

Install with a specific LLM provider:

pip install "amygdala[anthropic]"   # Anthropic (Claude)
pip install "amygdala[openai]"      # OpenAI (GPT)
pip install "amygdala[gemini]"      # Google Gemini
pip install "amygdala[ollama]"      # Ollama (local models)
pip install "amygdala[all-providers]"  # All providers

From source (development)

git clone https://github.com/sinanata/amygdala.git
cd amygdala
pip install -e ".[dev,all-providers]"

Quick Start

1. Initialize in your project

cd /path/to/your/project
amygdala init --provider anthropic --model claude-haiku-4-5-20251001

This creates a .amygdala/ directory with config and index files. Add .amygdala/ to your .gitignore if you don't want to track memory across machines.

Options:

• --provider -- LLM provider: anthropic, openai, gemini, ollama (default: anthropic)
• --model -- Model identifier (default: claude-haiku-4-5-20251001)
• --granularity -- Summary detail level: simple, medium, high (default: medium)
• --profile / -p -- Extension profile to enable (repeatable, see Extension Profiles)
• --auto-capture / --no-auto-capture -- MCP-driven auto-capture (default: enabled)

2. Capture file summaries

# Capture specific files
amygdala capture src/main.py src/utils.py

# Capture all tracked files
amygdala capture --all

# Capture with high granularity
amygdala capture --all --granularity high

3. Check status

amygdala status          # Rich table output
amygdala status --json   # JSON output (for scripting / hooks)

4. Detect dirty files

amygdala diff                          # Scan for files changed since last capture
amygdala diff --mark-dirty src/main.py # Manually mark a file as dirty

5. Install a platform adapter

amygdala install claude-code    # Install Claude Code hooks + MCP server
amygdala uninstall claude-code  # Remove adapter

Auto-Capture (MCP-Driven)

When auto_capture is enabled (the default), Amygdala keeps file summaries fresh using your Claude Code subscription -- no separate API key needed.

How it works

1. PostToolUse hook marks files dirty whenever Claude Code edits them
2. SessionStart hook injects project status (including dirty file list) into the session context
3. When Claude sees stale files in context, it uses the store_summary MCP tool to refresh them
4. Claude reads the file, writes a summary, and passes it to store_summary() -- the file is marked clean

The key insight: Claude Code is the LLM. Instead of making a separate API call to generate summaries, Claude generates them as part of its normal session activity.

Two capture modes

Mode	Who generates the summary	API key needed?	When to use
MCP-driven (default)	Claude Code itself	No	During active Claude Code sessions
CLI-driven (opt-in)	Anthropic / OpenAI / Ollama API	Yes	Batch capture, CI, non-Claude workflows

Disabling auto-capture

amygdala init --no-auto-capture

When disabled, the session context omits the auto-capture hint. You can still use amygdala capture with an API key, or manually call the MCP tools.

Extension Profiles

The base capture pipeline supports common source and config file types (.py, .js, .ts, .json, .yaml, etc.). Extension profiles add framework-specific file types, language mappings, and exclude patterns on top of this base set.

Enable profiles at init time with --profile / -p (repeatable):

# Single profile
amygdala init --profile unity

# Multiple profiles
amygdala init -p node -p react

# Check active profiles
amygdala config get profiles
amygdala status

Available Profiles

Profile	Key Extensions	Excludes
unity	.shader, .hlsl, .cginc, .compute, .unity, .prefab, .asset, .mat, .meta, .asmdef, .asmref, .shadergraph, .uxml, .uss	Library/, Temp/, Obj/, UserSettings/, Logs/
unreal	.inl, .uproject, .uplugin, .usf, .ush, .uasset, .umap	Binaries/, DerivedDataCache/, Intermediate/, Saved/
python	.pyi, .pyx, .pxd, .ipynb, .in, .conf	__pycache__/, .venv/, .tox/, .mypy_cache/, .pytest_cache/, .ruff_cache/
node	.mjs, .cjs, .mts, .cts, .npmrc	node_modules/, .next/, .nuxt/, .cache/, .turbo/
react	.scss, .sass, .less, .svg, .mdx	node_modules/, storybook-static/, coverage/
nextjs	.mdx, .scss, .svg	.next/, .vercel/, out/, node_modules/

Profiles compose via set-union -- enabling both node and react gives you all extensions from both. Language detection is also extended (e.g., .shader maps to shaderlab, .pyi maps to python).

Environment Variables

Set your API key for the provider you're using (only needed for CLI-driven capture):

Provider	Variable
Anthropic	ANTHROPIC_API_KEY
OpenAI	OPENAI_API_KEY
Gemini	GEMINI_API_KEY (or GOOGLE_API_KEY)
Ollama	(none -- local)

CLI Reference

Command	Description
amygdala init	Initialize Amygdala in a project
amygdala capture	Capture file summaries (specific or --all)
amygdala status	Show project memory status
amygdala diff	Scan for dirty files
amygdala config show	Show current configuration
amygdala config get	Get a config value (dot notation)
amygdala install	Install a platform adapter
amygdala uninstall	Remove a platform adapter
amygdala serve	Start the MCP server
amygdala clean	Remove all .amygdala/ data (--force)

Claude Code Integration

After installing the Claude Code adapter, Amygdala provides:

Hooks:

• SessionStart -- injects branch info, tracked/dirty file counts, and auto-capture hints into session context
• PostToolUse -- marks files dirty when Claude Code writes or edits them

MCP Server (7 tools available to Claude):

• get_file_summary(file_path) -- retrieve a file's cached summary
• get_project_overview() -- project-wide memory status
• list_dirty_files() -- files changed since last capture
• capture_file(file_path, granularity) -- capture using the configured LLM provider (requires API key)
• read_file_for_capture(file_path) -- read file content for summary generation
• store_summary(file_path, summary, granularity) -- store a summary you wrote (no API key needed)
• search_memory(query) -- search across all stored summaries

Start the MCP server:

amygdala serve

Project Structure

.amygdala/
  config.toml     # Provider, model, granularity settings
  index.json      # Per-file tracking (hash, status, timestamps)
  memory/         # Markdown summaries mirroring your source tree
    src/
      main.py.md
      utils.py.md

Granularity Levels

Level	Description
simple	One-line purpose of the file
medium	Purpose, key functions/classes, dependencies (default)
high	Detailed breakdown: every function, class, import, side-effect

Development

# Install dev dependencies
pip install -e ".[dev,all-providers]"

# Run tests
pytest

# Run with coverage
pytest --cov=amygdala --cov-report=term

# Lint
ruff check src/ tests/
ruff format src/ tests/

# Type check
mypy src/amygdala/

Requirements

• Python >= 3.12
• Git (project must be a git repository)

License

MIT