Cross-Agent Memory Bridge providing persistent memory layer for AI coding agents across sessions and IDEs, supporting 7 major agents including Cursor, Windsurf, and Claude Code.
Memorix is an MCP (Model Context Protocol) based persistent memory system designed to solve the "memory fragmentation" problem of AI coding agents.
Core Capabilities#
Intelligent Memory Storage
- Auto-categorizes into 9 observation types: session-request, gotcha, problem-solution, how-it-works, what-changed, discovery, why-it-exists, decision, trade-off
- 3-layer progressive disclosure search: L1 index search (~50-100 tokens) → L2 timeline context → L3 full details (~500-1000 tokens)
topicKeydeduplication mechanism for in-place updates, preventing memory redundancy
Knowledge Graph
- Compatible with MCP Official Memory Server API
- Supports entity extraction, relation inference, and observation addition
- Provides standard operations: create_entities, create_relations, search_nodes, read_graph
Cross-Agent Sync
memorix_workspace_sync: Scans source config, generates target Agent compatible config (merge without overwrite)memorix_rules_sync: Supports 6 format conversions: .mdc ↔ CLAUDE.md ↔ .kiro/steering/memorix_skills: Auto-clusters observations to generate SKILL.md documents
Implicit Memory Hooks
memorix hooks installone-click installation- Auto-detects Chinese/English decision patterns (e.g., "I decided to...", "The bug was...")
- 30-second cooldown period, intelligently skips trivial commands
Visual Dashboard
- Run
memorix_dashboardto launch Web UI (http://localhost:3210) - D3.js force-directed graph visualization
- Observation browser, memory decay panel, project switcher
- Light/dark themes, bilingual Chinese/English interface
Technical Implementation#
- Protocol: MCP Server via stdio communication
- Storage: 100% local at
~/.memorix/data/<projectId>/ - Search Engine: Orama full-text/vector search + JSONL/JSON persistence
- Optional Vector Enhancement: fastembed (native speed) or @huggingface/transformers (universal compatibility)
- Primary Language: TypeScript (87.1%), 422 unit tests coverage
- Runtime: Node.js ≥ 20.0.0
Typical Use Cases#
- Cross-session Memory: Architecture decisions discussed Monday (e.g., JWT + 15min expiry) auto-retrieved and applied in Tuesday's session
- Cross-Agent Collaboration: Payment module race condition fixed in Windsurf, Claude Code automatically knows and verifies
- Gotcha Prevention: Windows path separator issue recorded week 1, context auto-injected week 3 to prevent repeat errors
- IDE Migration: Migrate from Cursor to Kiro with one-click sync of MCP config + rules + skills
- Skill Accumulation: 50+ observations auto-cluster after 2 weeks, generating auth-module-guide.md, database-migrations.md, etc.
Quick Start#
# Recommended: Global install
npm install -g memorix
# MCP Config (universal for all agents)
{
"mcpServers": {
"memorix": {
"command": "memorix",
"args": ["serve"]
}
}
}
Note: Not recommended to run directly with npx as it may cause MCP initialization timeout (60-second limit).