Kraang Long-term memory for your agents.

Kraang is an MCP (Model Context Protocol) server that gives AI assistants persistent memory and session indexing, backed by SQLite with FTS5 full-text search and optional hybrid semantic search. It stores knowledge notes, indexes conversation transcripts, and surfaces what matters via search.

Why?

AI assistants forget everything between sessions. Kraang gives them persistent memory — decisions, debugging breakthroughs, patterns — so your next conversation picks up where the last one left off.

Requires Python 3.11+

Quick Start

The fastest way to get started is with kraang init:

uvx kraang init        # ephemeral — downloads on each run
uv tool install kraang # persistent — install once, use everywhere
kraang init

This creates a .kraang/ directory, initializes the database, configures .mcp.json, adds .kraang/ to .gitignore, sets up a SessionEnd hook for automatic session indexing, creates .claude/rules/kraang.md with agent usage guidelines, and indexes any existing sessions.

Manual Configuration

Add to your MCP client configuration (e.g. Claude Code, Claude Desktop):

{
  "mcpServers": {
    "kraang": {
      "command": "uvx",
      "args": ["kraang", "serve"],
      "env": { "KRAANG_DB_PATH": ".kraang/kraang.db" }
    }
  }
}

MCP Tools

Tool	Description
remember	Save a note. Updates in place if the title exists, otherwise creates a new one.
recall	Search notes and conversation sessions. Scope to "notes", "sessions", or "all".
read_session	Load a full conversation transcript by session ID (use recall to find sessions first).
forget	Downweight or hide a note by adjusting its relevance score (0.0 = hidden, 1.0 = full).
context	Auto-recall relevant memories and return safety-framed XML for context injection.
status	Get a knowledge base overview: note/session counts, recent activity, top tags.

CLI Commands

Command	Description
kraang init [PATH]	Set up kraang for the current project (database, .gitignore, config, hooks, rules, initial index).
kraang serve	Run the MCP server (invoked by Claude Code).
kraang index [PATH] [--from-hook]	Index or re-index conversation sessions for the project.
kraang sessions [-n LIMIT] [PATH]	List recent conversation sessions (default: 20).
kraang session <id> [-n MAX_TURNS]	View a session transcript in detail.
kraang search <query> [-n LIMIT]	Search notes and sessions (default: 10).
kraang notes [-a] [-n LIMIT]	List notes in the knowledge base (-a includes forgotten).
kraang status	Show knowledge base health and statistics.

Architecture

Kraang uses a layered architecture:

1. Models (models.py) -- Pydantic schemas for notes, sessions, and search results.
2. Store (store.py) -- SQLite backend with FTS5 full-text search, BM25 ranking, vector storage, and brute-force cosine fallback.
3. Search (search.py) -- Query parsing and FTS5 expression building.
4. Hybrid (hybrid.py) -- Hybrid vector + keyword search with weighted merge and FTS fallback.
5. Embeddings (embeddings.py) -- Embedding provider abstraction with OpenAI and local (fastembed) implementations, retry logic, and L2 normalization.
6. Safety (safety.py) -- Prompt injection detection and safe context formatting for LLM consumption.
7. Indexer (indexer.py) -- Reads Claude Code JSONL transcripts and indexes sessions.
8. Server (server.py) -- MCP server exposing 6 tools over stdio.
9. CLI (cli.py) -- Typer CLI for init, serve, index, and local queries.
10. Formatter (formatter.py) -- Markdown formatting for MCP tool responses.
11. Display (display.py) -- Rich console rendering for CLI commands.
12. Config (config.py) -- Project root detection, database path resolution, and title normalization.

Semantic Search (optional)

Kraang supports hybrid semantic + keyword search for more intelligent recall. Queries like "that debugging trick from last week" work alongside exact keyword matches.

Semantic search works out of the box using local embeddings via fastembed (ONNX-based, runs on CPU). No API key required.

Providers

Provider	Model	Dims	Setup
Local (default)	nomic-ai/nomic-embed-text-v1.5-Q	768	Works automatically
OpenAI	text-embedding-3-small	1536	Set OPENAI_API_KEY

How it works

• Hybrid search: Combines vector similarity (70%) with keyword matching (30%)
• Zero config: Local embeddings are used by default — no API key needed
• Graceful degradation: If both providers fail, keyword search still works
• Embedding cache: Content-addressed cache avoids redundant computation/API calls
• Safety framing: The context tool wraps recalled memories with prompt injection protection

Environment variables

Variable	Required	Default	Purpose
KRAANG_EMBEDDING_PROVIDER	No	auto-detect	Force "openai" or "local"
KRAANG_EMBEDDING_MODEL	No	—	Override the local model name
OPENAI_API_KEY	No	—	Enables OpenAI embeddings (auto-detected when set)
KRAANG_DB_PATH	No	<project_root>/.kraang/kraang.db	Override database location

Claude Code integration

When using kraang init, optionally add OPENAI_API_KEY to the generated .mcp.json to use OpenAI embeddings instead of local:

{
  "mcpServers": {
    "kraang": {
      "command": "uvx",
      "args": ["kraang", "serve"],
      "env": {
        "KRAANG_DB_PATH": ".kraang/kraang.db",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Development

git clone https://github.com/johnnygreco/kraang.git && cd kraang
uv sync --extra dev
make install-hooks   # install pre-commit hooks (run once)
make test
make lint

Pre-commit hooks run automatically before each commit (ruff format, ruff check --fix, ty). Run manually:

uv run pre-commit run --all-files

Run the full check suite:

make coverage   # tests + coverage report
make format     # auto-format with ruff

Troubleshooting

Problem	Fix
Kraang tools not showing up in Claude Code	Restart Claude Code after running kraang init
Sessions not being indexed automatically	Check that .claude/settings.json has the SessionEnd hook
Search returns nothing	Run kraang status to check counts, then kraang index to re-index
Need a fresh start	Delete .kraang/ and re-run kraang init

License

Apache 2.0