Stop Managing Information. Start BUILDING Intelligence.
Working Mind is a terminal AI agent with a persistent knowledge graph. Not a chatbot that searches old messages -- a cognitive layer that architects understanding. Entities, relations, observations -- structured, connected, and compounding every session. Your intelligence. Your machine. Your graph.
Three Layers. One Cognitive Engine.
Working Mind is not a single model hoping to guess right. It is the integration of three inseparable layers -- a reasoning core that interprets intent, a graph memory that stores structured knowledge, and a modular toolkit that acts in the world. Each layer is necessary. Together, they are sufficient.
- The LLM reads your ambiguous goal, breaks it into hypotheses, and directs the workflow. It reasons about what to look up, what to connect, and what to verify -- it does not just generate text.
- Entities, Relations, Observations -- not chat logs, not documents, not keyword indexes. Every piece of knowledge is context-mapped. No insight is truly lost; it is stored relationally, queryable by meaning, not by scroll.
- Packs wire the agent for your domain. Search the web, scrape documents, query APIs, read local files. Extensible via MCP. Not a generic skill that self-improves by guessing -- a focused configuration you control.
It Compounds. Not Just Persists.
Chat logs persist but they do not compound. After ten sessions with a chatbot, you have ten chat logs. After ten sessions with Working Mind, you have a knowledge graph that answers questions no single session could. Cross-links emerge. Patterns surface. Contradictions are caught. Your expertise deepens -- because the graph compounds.
- People, projects, compounds, cases, statutes, technologies. Each gets its own node with observations attached. Search by meaning, not by scrolling through chat history.
- "Taaffeite --confused-with--> Spinel." The agent connects entities when you mention them together. Connections you forgot you made -- patterns no keyword search surfaces.
- When you store "prefers Python" today and "prefers Rust" tomorrow, Working Mind detects the conflict, invalidates the old observation, and preserves the timeline. The graph corrects itself -- with your approval.
You Curate. It Remembers.
Other agents auto-save everything and hope the search finds it later. Working Mind only stores what you approve. You teach it, you verify the answers, you correct the graph. When the graph is right, it is because you made it right. That is what makes it research-grade.
- "Remember that taaffeite has refractive index 1.719-1.730." Working Mind saves entities, relations, and observations to your local knowledge graph -- structured, not buried in chat logs.
- Ask a question. If the answer is wrong, correct the graph. Ask again. When the answer is right, the graph is right. You are the quality gate -- every fact is one you approved.
- Working Mind does not nudge itself to remember things. It does not auto-create skills from experience. It does not profile you. You decide what goes into the graph. Nothing else does.
Secure by Constraint. Private by Design.
Working Mind is not a cloud service. It is a terminal application that runs on your machine. No inbound ports. No telemetry. No accounts. No messaging gateways. Your data never leaves your disk except as API calls to the LLM provider you chose.
- No web dashboard. No database to breach. No login. MCP tools run as local child processes over stdio -- no HTTP, no sockets, no ports. Your data never leaves your machine except as LLM API calls.
- OpenAI, Anthropic, Google, Ollama, OpenRouter, or any OpenAI-compatible endpoint. API keys from environment variables only. Never stored in plaintext. Never sent anywhere except your chosen provider.
- All inference, all tools, all storage on your machine. MCP tools run as local processes over stdio. No cloud runtime. No gateway process. No VPS. Your data never leaves your disk except as LLM API calls to the provider you chose.
- Knowledge graph is an SQLite database you can query. Config is a JSON file you can edit with vim. Packs are directories of markdown. A working system you can understand end to end in an afternoon.
Built for Research-Grade Work
No telemetry. No accounts. No cloud. No autonomous memory. You install it, you run it, you own the data.
Knowledge Graph
Structured entities, relations, and observations -- not keyword search over old messages. Your agent queries the graph before every answer. Contradiction detection keeps it honest.
User-Curated
You decide what goes in the graph. No auto-saving. No nudges. No user profiling. When the graph is right, it is because you made it right.
7 Providers, 75+ Models
Local Fast, Ollama, OpenRouter, Together, Google, OpenAI, Groq. Plus Anthropic and DeepSeek via OpenRouter. You bring your own key. You choose your model.
MIT Licensed
Audit the code. Fork it. Build your own packs. No gatekeeper. The entire source is on GitHub.
MCP Ecosystem
Connect Brave Search, Firecrawl, GitHub, Postgres, or any MCP server. Tools appear immediately. You approve every call.
Temporal Validity
Every observation has a timeline. Old facts are invalidated, not deleted. Search the graph as it was last month, or as it is now. History is preserved.
Start building your permanent knowledge graph
Install Working Mind. Teach it what you know. Verify the answers. Watch your expertise compound across sessions.