AgenticMind

What is AgenticMind?

AgenticMind is a self-hosted, auditable knowledge and memory layer for AI agents, served over the Model Context Protocol (MCP). It provides citation-enforced answers with a full why-trace, a self-improving corpus, and runs entirely on Postgres with pgvector—no additional vector stores or external dependencies required for embeddings.

How to use AgenticMind?

Quickstart: run OPENAI_API_KEY=sk-... sh -c "$(curl -fsSL https://raw.githubusercontent.com/Moai-Team-LLC/AgenticMind/main/quickstart.sh)" to pull Docker images and bring up the server. From source, clone the repo, run ./setup.sh then npm run dev. Point any MCP client (Claude Code, Cursor, LangGraph) at http://localhost:3000/mcp with the auto-generated bearer token.

Key features of AgenticMind

• Citation-enforced answers: every claim tied to a numbered source.
• Replayable why-trace per answer: retrieved, ranked, and used.
• Self-improving corpus via judge-gated compounding loop.
• Tiered retrieval: chunks, typed fact cards, knowledge graph.
• Single datastore: Postgres + pgvector for all indexes.
• Zero-key offline embeddings (bge-m3) for multilingual retrieval.

Use cases of AgenticMind

• Agents that must answer from trusted sources with provable citations.
• Systems requiring a replayable audit trail for every answer.
• Self-hosted knowledge bases that improve automatically without human feedback.
• Multi-agent setups that need governed, accountable memory across agents.
• Production deployments where a single Postgres database replaces vector-store sprawl.

FAQ from AgenticMind

How does AgenticMind differ from plain RAG or memory SDKs?

AgenticMind enforces citations, provides a full why-trace for every answer, and includes a self-improving corpus and relational verification via a knowledge graph—all on Postgres+pgvector, unlike most RAG solutions that offer only fuzzy recall.

What are the runtime and database requirements?

Requires Node ≥22.18 or Bun ≥1.3, and PostgreSQL with pgvector. Embeddings run locally by default (bge-m3), so no cloud key is needed for retrieval; only a chat-model key (e.g., OpenAI API key) is required for answer synthesis.

How does the corpus self-improve?

Validated answers are promoted back into the corpus through a judge-gated compounding loop driven by programmatic signals—not human feedback. A kl_signal tool emits these signals on prior answers.

Is there a frontend or UI?

No. AgenticMind is headless; agents interact with it solely via MCP tools over streamable HTTP. The only consumers are agents.

What authentication does the MCP server use?

The MCP endpoint uses per-token bearer authentication with scoped, least-privilege tokens. In the quickstart, a static bearer key (MCP_API_KEY) is auto-generated; from source, a JWT with typ="mcp" is accepted. Authorization is fail-closed.