omega-memory

OMEGA is a persistent memory system designed for AI coding agents, emphasizing local-first and cross-model compatibility. It aims to solve the problems of AI coding agents being stateless, starting from scratch in every new session, repeating mistakes, losing context, and the vendor lock-in and data privacy issues brought by cloud memory services.

Purpose

OMEGA provides persistent memory, coordination, and learning capabilities for AI coding agents, enabling agents to remember decisions, preferences, architectural choices, error patterns, and lessons learned across sessions, thereby reducing the time spent repeatedly explaining context and avoiding repeated mistakes. It is suitable for developers using mainstream LLMs and coding agents such as Claude, GPT, Gemini, Cursor, and Claw Code, and is especially suitable for complex projects that need to maintain contextual consistency across sessions.

Main Capabilities

• Memory and Learning: Stores decisions, lessons learned, error patterns, and preferences, and supports semantic search. Tools cover compression, consolidation, timelines, graph traversal, and checkpoint/restore.
• Multi-Agent Coordination (omega-pro): File and branch locking, session management, dependency-aware task queues, intent broadcasting, and inter-agent messaging, preventing agents from overwriting each other's work.
• Intelligent LLM Routing (omega-pro): Classifies tasks and routes them to the optimal model, for example routing coding tasks to Claude Sonnet, quick edits to Llama 8b (at 1/60 the cost), and 1 million token contexts to Gemini Flash. Supports 5 providers and 4 priority modes, with intent classification taking less than 2 milliseconds.
• Knowledge Base (omega-pro): Ingests PDFs, Markdown, web pages, and text files into a searchable knowledge base, with support for semantic chunking.
• Entity Registry (omega-pro): Multi-entity enterprise memory, supporting relationships, hierarchies, and entity-scoped memories/profiles/documents.
• Secure Profile (omega-pro): Personal data storage encrypted with AES-256, integrated with macOS Keychain.

Usage

OMEGA requires Python 3.11 or higher. When installing, it is recommended to use an isolated environment, such as installing via pipx or a virtual environment omega-memory[server](note that quotes must be preserved in zsh). After installation, run omega setup to configure; this command creates the ~/.omega/ directory, downloads the ONNX embedding model (bge-small-en-v1.5, about 130 MB) and the reranking model (about 90 MB) to ~/.cache/omega/models/, registers the MCP server, installs session hooks, and modifies ~/.claude/CLAUDE.md.

OMEGA works through natural language, with no API calls or configuration required. Users can directly tell Claude to remember something, for example, "Remember that the authentication system uses JWT tokens, not session cookies," and Claude will store it as a permanent memory with semantic embeddings. After closing a session and opening a new one, when asking related questions, OMEGA will automatically surface the relevant memories.

OMEGA runs as an MCP server within Claude Code. By default, it exposes five tools to the client (omega_store, omega_welcome, omega_protocol, omega_tools, omega_call) to save context. Set the environment variable OMEGA_CONDENSED=0 Can expose all tools. OMEGA Pro adds coordination, routing, entity, knowledge base, and profile tools.

Use cases

• AI programming projects that need to maintain consistent context across sessions.
• Developers who use multiple LLMs or programming agents and want to avoid vendor lock-in.
• Teams that need multi-agent collaboration, task queues, and file locking.
• Scenarios that need to ingest materials such as documents and web pages into a knowledge base for semantic search.
• Users with data privacy requirements who want memory to be stored entirely locally.

Limitations

• Requires Python 3.11 or later; the Python 3.9 bundled with macOS is too old.
• The free version has storage limits: when storage reaches 2,000 memories, search switches from semantic search to plain text matching; at 5,000, it stops storing new memories. OMEGA Pro removes these two limits.
• Hooks (automatic capture and presentation) apply only to Claude Code.
• When installed only as a Python library, there is no automatic capture or presentation functionality, because hooks require an MCP server.
• Some advanced features (multi-agent coordination, LLM routing, knowledge base, entity registration, security profiles) require OMEGA Pro.
• Installation and configuration may modify ~/.claude.json、~/.claude/settings.json and ~/.claude/CLAUDE.md and other files.
• Official materials do not provide specific pricing information.