Introduction

HUSK is a self-hosted memory layer for AI coding assistants. It captures what you work on, remembers cross-project patterns, and surfaces relevant context — across all your machines and tools.

Key features

• Semantic search — find memories by meaning, not just keywords. Multiple embedding backends: local (Ollama, transformers, llama.cpp) or cloud (OpenAI, Voyage).
• Knowledge graph — link memories with typed edges (caused_by, contradicts, supersedes, related_to), traverse causal chains, and find contradiction clusters.
• Memory scopes — session, project, and global memories with automatic TTL. Keep the right context at the right level.
• Client-agnostic — works with any MCP-compatible client. Claude Code plugin included, but the server doesn't care what sends the data.
• Duplicate detection — automatically detects and prevents storing near-identical memories using configurable similarity thresholds.
• Session tracking — records what you work on across sessions, summarises past work, and lets you pick up where you left off.
• Self-hosted — your data stays on your machines. Run via the CLI or Docker Compose.

How to get started

1. CLI — install and configure HUSK with npx @huskai/cli
2. Quick start — or get running with Docker Compose in under a minute
3. Connecting — wire up Claude Code or any MCP client
4. Plugins — session hooks and skills beyond raw MCP
5. MCP Tools — reference for all available tools

How does HUSK compare?

See how HUSK stacks up against other AI memory tools at ai-memory-comparison.vercel.app.