Skip to main content

How OMEGA Works

A single MCP server that connects your AI client to a local SQLite database with vector search and encryption.

MCP Client
Claude Code, Cursor, Windsurf
OMEGA MCP Server
12 tools · Hook integration
Local Storage
SQLite + sqlite-vec
ONNX Runtime (CPU)
macOS Keychain

12 Tools, Five Capabilities

One memory system with semantic search, auto-capture, learning, forgetting, and graph relationships.

Beyond Storage

OMEGA doesn't just store and retrieve. It re-ranks results for precision and detects when new information contradicts what it already knows.

Cross-Encoder Reranking

Initial retrieval uses fast vector similarity to find candidates. Then a neural cross-encoder (ms-marco-MiniLM-L-6-v2) re-scores the top 20 results by encoding each query-passage pair together, catching semantic relationships that separate embeddings miss.

Modelms-marco-MiniLM-L-6-v2 (ONNX, 22 MB)
CandidatesTop 20 rescored per query
SpeedCPU-only, no GPU required
FallbackGraceful. Standard retrieval if unavailable
Contradiction Detection

When you store a new memory, OMEGA checks it against existing ones. Four heuristic signals detect conflicts: negation asymmetry, antonym pairs, preference value changes, and temporal overrides. Both memories get annotated with bidirectional metadata.

SignalsNegation, antonyms, preferences, temporal
GateCross-encoder similarity pre-filter
OutputBidirectional metadata + graph edges
CostZero. Pure heuristics, no LLM calls

Full Comparison

Honest, side-by-side. Tool counts are approximate and based on publicly available documentation as of Feb 2026.

FeatureOMEGAMem0ZepLettaNative
MCP Tools1299-1070
Semantic SearchYesYesYesYesNo
Auto-Capture & SurfacingYesYesCloud onlyNoPartial
Cross-Session LearningYesYesYesYesLimited
Intelligent ForgettingYesNoNoNoNo
Cross-Encoder RerankingYesNoNoNoNo
Contradiction DetectionYesNoNoNoNo
Graph RelationshipsYesPro onlyYesNoNo
Checkpoint / ResumeYesNoNoNoNo
Encryption at RestAES-256EnterpriseCloudNoNo
Local-FirstYesNoNoYesYes
No External DB RequiredSQLiteCloudNeo4jPostgresFile
LongMemEval Score95.4%N/A71.2%N/AN/A

Benchmark Breakdown

LongMemEval (ICLR 2025) scores by category. 500 questions across 5 capability areas.

Temporal reasoning is a known weakness. I publish honest numbers because trust matters more than optics.

Single-Session Recall125 / 126
99%
Preference Application30 / 30
100%
Multi-Session Reasoning111 / 133
83%
Knowledge Updates75 / 78
96%
Temporal Reasoning125 / 133
94%

Category scores from our 95.4% task-averaged accuracy (466/500 raw). Methodology: LongMemEval.

Real Numbers

~31MB
Startup
~337MB
After First Query
~50ms
Retrieval
SQLite
Storage
None
GPU

Measured on M1 MacBook Pro, ~240 memories, bge-small-en-v1.5 ONNX model. RSS via Activity Monitor.

10,000+
MCP servers
97M
monthly SDK downloads
Linux Foundation
governance

Built for MCP

The Model Context Protocol is the fastest-adopted standard in AI tooling. OMEGA provides the persistent memory infrastructure that MCP doesn't standardize yet.

Works with any MCP client. No vendor lock-in. Ecosystem stats from Anthropic and community registries, Q1 2026.

What You Need

Minimal dependencies. Runs on hardware you already have.

Platform
  • Python 3.11+
  • macOS / Linux
  • No GPU required
MCP Clients
  • Claude Code
  • Cursor
  • Windsurf
  • Any MCP-compatible client
Storage
  • SQLite (built-in)
  • ~337 MB RAM after first query
  • ~10 MB per 250 memories