AI assistants forget everything between conversations. Existing solutions — vector databases, RAG pipelines, memory plugins — all share the same problems:
- Invisible — memories are buried in databases you can't see or inspect
- Uncontrollable — no way to edit, merge, or delete individual memories
- Messy — duplicates, contradictions, and outdated info pile up with no cleanup
- Static — the system only knows what you explicitly tell it
MindReader is a personal knowledge graph that gives you full visibility and control over your AI's memory.
See your entire memory landscape as an interactive graph. Nodes are entities (people, projects, tools), edges are relationships. Zoom, filter, click to explore. No more black boxes.
Click any entity to edit its summary, tags, and category. Merge duplicates. Create or delete relationships. Preview blast radius before deleting. Your memory, your rules.
Select any node and click Evolve. MindReader sends it to a web-search-capable LLM that researches the entity, discovers related people/projects/concepts, and builds relationships — all streamed live onto the graph.
self-evolving-upload-hq.mp4
A single evolve typically discovers 10-25 new entities and relationships, turning an isolated node into a rich knowledge cluster.
Memories naturally decay over time unless reinforced by frequent use. Every entity and relationship has a strength score that decays exponentially — frequently accessed memories stay vivid, unused ones fade and eventually expire.
- Visual decay — node opacity reflects memory strength, faded nodes indicate aging knowledge
- "Show Decay" mode — color the graph by strength (green → yellow → red)
- Time Travel — drag a slider to see the graph at any point in time. Nodes that didn't exist yet disappear, expired ones reappear as ghosts. Hit Auto Play to watch your knowledge graph grow from scratch.
- Restore — one-click restore any expired entity
LLM-powered background maintenance keeps the graph clean: auto-categorization, auto-tagging, duplicate detection, relationship repair, orphan cleanup — no manual intervention needed.
Browse your knowledge as a list, timeline, graph, categories, or activity log.
MindReader is a first-class extension for OpenClaw. Install once, then every conversation automatically builds your knowledge graph.
- Auto-capture — silently extracts entities, facts, and relationships from conversations
- Auto-recall — before the agent responds, retrieves relevant memories and injects them as context
- Tool calls — agents can explicitly search, store, and list memories
- Web UI — access the full graph explorer anytime at
http://localhost:18900
Just talk to your agent naturally. MindReader handles the rest.
MindReader includes an MCP server for integration with Claude Code, Cursor, or any MCP-compatible client. Six tools: search, store, create, recall, list entities, and stats.
{
"mcpServers": {
"mindreader": {
"command": "node",
"args": ["/path/to/mindreader-mcp/src/index.js"]
}
}
}git clone https://github.com/flu012/mindreaderv2.git
cd mindreaderv2
npm run setup # interactive wizard
npm start # http://localhost:18900- Node.js 18+
- Python 3.11+
- Docker (for Neo4j database)
- LLM API key (any of the providers below)
| Provider | Default Model | Web Search | Notes |
|---|---|---|---|
| OpenAI | gpt-4o-mini |
No | Most widely available |
| DashScope (Alibaba) | qwen3.5-flash |
Yes (built-in) | Best choice for self-evolve |
| Anthropic | claude-sonnet-4-6 |
No | Native API support |
| Ollama | llama3.2 |
No | Free, fully local, no API key needed |
Self-evolve works significantly better with models that support built-in web search (e.g. DashScope's Qwen series). Ollama is ideal for privacy-first setups where all data stays local.
For systems that need precise, deterministic memory management without LLM processing:
curl -X POST http://localhost:18900/api/entities -H "Content-Type: application/json" \
-d '{"entities":[{"name":"Alice","category":"person","tags":["engineer"]}]}'Supports batch creation, upserts, and relationship management. See the API reference.
- Quick Start Guide — step-by-step install, provider setup, OpenClaw integration, troubleshooting
- Technical Reference — architecture, preprocessing pipeline, configuration, CLI reference, development
MindReader V2 is an experimental project for personal use and research. Provided as-is with no guarantees of data integrity, security, or availability. Memory data is stored in local Neo4j — back up your database regularly. LLM-generated content (auto-categorization, evolve results, extracted facts) may contain inaccuracies — review important memories manually. API keys and credentials in .env are stored in plaintext — do not commit this file to version control.
MIT