OMEGA is a local-first memory layer for AI coding agents, so decisions, preferences, lessons, and error patterns survive from one session to the next. Memories live in a single SQLite database on your machine, embeddings are computed on-device with an ONNX model, and nothing goes to a cloud service. Because the memory sits outside any one model provider, it carries across Claude, GPT, Gemini, and other models.
It runs as an MCP server that any MCP client can use, with setup for Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Codex, and Antigravity, and it also works as a plain Python library for scripts and CI. In Claude Code, hooks capture decisions and lessons automatically, give a briefing at session start, and surface relevant memories after edits and reads. To save context, the server exposes five tools by default and routes the rest through a single call tool.
Retrieval combines vector similarity, FTS5 keyword search, type weighting that favors decisions and lessons, contextual re-ranking, and deduplication, and returns a small ranked slice of context rather than a full memory dump. Memories deduplicate on write, merge new insights into similar entries, link to related memories, and compact into summaries over time, and session summaries expire after a day while lessons and preferences persist. OMEGA reports 95.4% on the LongMemEval benchmark in a research setup with GPT-4.1 as the reader model.
| Feature | OMEGA | Anthropic Memory | Mem0 | Zep |
|---|---|---|---|---|
| Works with any LLM or agent | Yes | Claude only | Yes | Yes |
| Data stays on your machine | Yes | Partial | No | No |
| No cloud dependency | Yes | No | No | No |
| Semantic search and knowledge graph | Yes | No | Paid tier | Yes |
The core is free, with semantic search falling back to text match past 2,000 memories and storage capped at 5,000. OMEGA Pro removes those limits and adds multi-agent coordination (file and branch locks, task queues, agent messaging), task-based routing across model providers, a document knowledge base, an entity registry, and encrypted profile storage.
Features
- Typed memories: store decisions, lessons, errors, and preferences with semantic embeddings
- Hybrid retrieval: vector similarity, full-text search, type weighting, and re-ranking return a small ranked context slice
- Automatic capture: Claude Code hooks record decisions and surface relevant memories without explicit prompts
- Checkpoint and resume: save task state and pick it up in a later session
- Memory maintenance: deduplication, consolidation, compaction, contradiction finding, backup, and restore
- Reminders and profile: time-based reminders and a stored user profile with preferences
- Code review with memory: review a diff against your codebase's remembered conventions
- Second opinions: consult GPT or Claude from inside a session with your own key
- Client setup: registers with Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Codex, Antigravity, or any MCP client
- Python library: the full storage and retrieval API without running an MCP server
- Pro coordination: file and branch locking, task queues, and agent-to-agent messaging for parallel sessions
- Pro routing: classifies tasks and routes them across five model providers
