/omega-memory

name: omega-memory
description: "Persistent memory for AI coding agents. Teaches agents how to use OMEGA's MCP tools for storing decisions, querying context, coordinating multi-agent workflows, and resuming tasks across sessions."
license: Apache-2.0
compatibility: "Python 3.11+, Claude Code, Cursor, Windsurf, Zed"
metadata:
  category: memory
  pypi: omega-memory
  github: omega-memory/omega-memory

OMEGA Memory

Persistent memory for AI coding agents. OMEGA gives your agent a knowledge graph it can query, learn from, and coordinate through across sessions.

This skill teaches you how to use OMEGA's MCP tools effectively.

Setup

pip3 install omega-memory[server]
omega setup        # auto-configures your editor + downloads embedding model
omega doctor       # verify everything works

Works with Claude Code, Cursor, Windsurf, Zed, and any MCP client.

Core Tools

OMEGA provides 12 MCP tools. Here's when to use each one.

Storing Memories

**`omega_store(content, event_type, metadata?, entity_id?)`**

Store decisions, lessons, and context that should persist across sessions.

| Event Type | When to Use | TTL | |------------|------------|-----| | `decision` | Architectural choices, technology selections | 90 days | | `lesson_learned` | Debugging insights, patterns that worked/failed | 90 days | | `user_preference` | Code style, workflow preferences, tool choices | Permanent | | `error_pattern` | Recurring errors and their fixes | 30 days | | `task_completion` | Completed work with outcomes | 14 days | | `checkpoint` | Mid-task state for resumption | 7 days |

omega_store("Switched from REST to GraphQL for the dashboard API — reduces N+1 queries", "decision")
omega_store("User prefers early returns, max 2 levels of nesting", "user_preference")
omega_store("pytest fixtures with db cleanup must use function scope, not session scope", "lesson_learned")

**Don't store:** Raw code output, tool results, transient status updates, anything shorter than a sentence.

Querying Memories

**`omega_query(query, mode?, limit?, entity_id?)`**

Search memories by meaning, not just keywords. Uses hybrid retrieval: vector similarity + full-text search + cross-encoder reranking.

| Mode | When to Use | |------|------------| | `semantic` (default) | Find memories by meaning — "how did we handle auth?" | | `phrase` | Exact substring match — find a specific term or identifier | | `timeline` | Recent memories grouped by day — "what happened this week?" | | `browse` | List by type, session, or recency — explore what's stored |

omega_query("database migration strategy")
omega_query("what decisions were made about the API", mode="timeline", days=7)
omega_query("pytest", mode="phrase")
omega_query(mode="browse", browse_by="type")

**Pro tip:** Query before starting work. Prior decisions and lessons save time and prevent repeating mistakes.

Session Management

**`omega_welcome(project?)`** — Call at session start. Returns recent context, active reminders, and project state. This is how your agent picks up where it left off.

**`omega_checkpoint()`** — Save current task state mid-session. If the session ends unexpectedly, the next `omega_welcome` restores this context.

**`omega_resume_task(task_id)`** — Resume a previously checkpointed task with full context.

Memory Maintenance

**`omega_reflect()`** — Analyze memory quality: duplicates, contradictions, coverage gaps.

**`omega_maintain(action)`** — Run maintenance operations: consolidation, compaction, health checks.

Retrieval Architecture

OMEGA's query pipeline runs 7 phases to find the most relevant memories:

1. **Vector similarity** — Embedding search (bge-small-en-v1.5, 384-dim) via sqlite-vec 2. **Full-text search** — FTS5 with BM25 scoring 3. **Strong signal short-circuit** — Skip expensive phases when FTS5 finds an exact match 4. **Score fusion** — Reciprocal Rank Fusion combines vector + text scores 5. **Contextual boosting** — Boost results matching current file, project, or tags 6. **Cross-encoder reranking** — ms-marco-MiniLM-L-6-v2 rescores top candidates 7. **Assembly** — Dedup, normalize, apply minimum relevance threshold

This hybrid approach achieves 95.4% on LongMemEval (500-question benchmark).

Best Practices

What to Store

• Architectural decisions with reasoning ("chose X because Y")
• Debugging insights that took effort to discover
• User preferences stated explicitly ("always use..." / "never...")
• Cross-session context that future sessions need

What NOT to Store

• Information already in the codebase (read the code instead)
• Transient state (build output, test results)
• Anything shorter than a meaningful sentence
• Speculative conclusions from reading a single file

Query Patterns That Work

• **Before starting a task:** `omega_query("prior decisions about [feature area]")`
• **Before modifying a file:** `omega_query(context_file="/path/to/file.py")`
• **After debugging:** `omega_store("[root cause and fix]", "lesson_learned")`
• **When user says "remember":** `omega_store("[what they said]", "user_preference")`

Anti-Patterns

| Don't | Do Instead | |-------|-----------| | Store every tool result | Store only insights and decisions | | Query with single words | Use natural language questions | | Skip `omega_welcome` at session start | Always call it — it loads critical context | | Store without `event_type` | Always specify type for proper TTL and dedup | | Guess from stale memory | Query OMEGA to verify current state |

How It Works Under the Hood

• **Storage:** SQLite with WAL mode. Single file at `~/.omega/omega.db`.
• **Embeddings:** bge-small-en-v1.5 via ONNX Runtime (~90MB RAM). LRU cache (512 entries).
• **Vector search:** sqlite-vec extension for ANN similarity search.
• **Text search:** FTS5 with BM25 ranking.
• **Dedup:** Jaccard similarity with per-type thresholds (0.70-0.90). Content-level and embedding-level.
• **Memory evolution:** Si