OmniMem
Provide persistent, self-hosted semantic memory and knowledge recall for AI agents across sessions and projects.
OmniMem is a self-hosted semantic memory system built in Rust on top of SQLite, serving memory tools to AI agents via the Model Context Protocol (MCP). It solves context loss and repetitive debugging by tracking episodic events, project context, user preferences, RSS-fed knowledge, and compiled skill documents. Featuring local ONNX embeddings, a dead-end 'graveyard' to prevent repeat mistakes, trust and contradiction detection, and experience-weighted scoring, OmniMem packages an MCP server, local vector store, and background enrichment pipeline into a single binary available as a desktop tray app or headless server.
Key Features
01Experience-weighted ranking algorithm combining vector similarity, recency, lifecycle state, and trust scores
02103 GitHub stars
03Multi-tier memory architecture covering episodic memories, project context, user preferences, skills, and RSS articles
04Single-binary architecture using SQLite and local ONNX Runtime embeddings without requiring external API calls
05Automated maintenance including semantic deduplication, contradiction detection, and skill compilation
06Dead-end graveyard and Claude Code integration to catch and halt repeated mistakes before tool execution
Use Cases
01Preventing coding assistants like Claude Code and Cursor from repeating known failed approaches or dead ends
02Sharing persistent context, preferences, and project architecture across disparate AI agents and client machines
03Automatically ingesting and retrieving domain-specific articles and documentation via scheduled RSS feeds