iikarus/Dragon-Brain — Dragon Brain — persistent long-term memory for AI agents via MCP (Model Context Protocol).

Description and use

**Brief Summary (2‑3 sentences)**
Dragon‑Brain is an open‑source framework that gives AI agents persistent long‑term memory by combining a graph database (FalkorDB) with vector search (Qdrant) and GPU‑accelerated embeddings. It implements the Model Context Protocol (MCP) and ships ready‑to‑use integrations for Claude, Gemini CLI, Cursor, Windsurf, VS Code Copilot and 30 other tools, backed by a comprehensive test suite (1,121 tests). With a score of 82/100, it is positioned as a production‑grade OSS candidate for building repeatable, memory‑aware agent workflows.

---

Value Proposition

• **Unified Memory Layer:** By persisting both relational knowledge (graph) and semantic embeddings, agents can recall past interactions, share context across sessions, and avoid “prompt forgetting.”
• **Plug‑and‑Play MCP Compliance:** The Model Context Protocol standardizes how prompts, tool calls, and memory updates are exchanged, making it easy to swap or combine LLM back‑ends (Claude, Gemini, etc.).
• **Tool‑Centric Orchestration:** Pre‑built connectors for 30 tools let developers quickly assemble multi‑agent pipelines (e.g., code generation → linting → testing) without writing custom glue code.
• **Scalable Performance:** CUDA‑accelerated embedding generation and Qdrant’s ANN search keep latency low even as the knowledge graph grows.

Practical Adoption Path

1. **Evaluate the API/SDK:** Clone the repo, run the provided Docker compose (or local Python environment) and interact with the REST/CLI endpoints to store and retrieve a few test facts. The test suite can be used as a sanity‑check.
2. **Integrate a Target LLM:** Replace the default Claude/Gemini client with your own model by implementing the MCP interface; the codebase already abstracts the model behind a thin wrapper.
3. **Connect Existing Tools:** Use the supplied adapters (e.g., cursor_adapter.py, copilot_adapter.py) as templates to wrap any internal tool or service you need the agent to call.
4. **Pilot a Workflow:** Define a simple multi‑step use case (e.g., “fetch data → summarize → generate report”) and let the agent store intermediate results in the graph. Observe how memory improves downstream steps.
5. **Scale & Harden:** Deploy FalkorDB and Qdrant in a managed Kubernetes cluster, enable GPU‑enabled embedding workers, and add monitoring/alerting around the MCP endpoints.

Production Readiness

• **Activity & Community:** Recent commits (as of 2026‑06‑24), 50 ★, 7 forks, and 13 GitHub topics indicate an active, albeit small, community. The extensive test coverage (1,121 tests) shows a focus on reliability.
• **Architecture:** Decoupled services (graph DB, vector store, embedding workers) align with cloud‑native deployment patterns; GPU support is optional but available for high‑throughput scenarios.
• **Ecosystem Fit:** Direct integrations with popular developer tools (VS Code Copilot, Cursor) and LLM providers make it easy to embed in existing pipelines.
• **Risks to Address Before Full Production:** A final review of the license (MIT‑like but verify), security posture of the exposed APIs, and confirmation of an active maintainer or governance model are still required.

Overall, Dragon‑Brain is mature enough for a serious pilot in environments that need persistent, context‑rich AI agents, with a clear upgrade path to full production once the remaining compliance checks are completed.

Problem

iikarus/Dragon-Brain helps turn isolated prompts and tools into repeatable agent workflows.

Outcome

High for an OSS candidate: recent activity, adoption, and ecosystem signals are strong enough for a serious pilot.

Workflow

• coordinate multi-agent workflows
• add tool-use pipelines
• standardize agent memory

Evidence

Dragon Brain — persistent long-term memory for AI agents via MCP (Model Context Protocol). Knowledge graph (FalkorDB) + vector search (Qdrant) + CUDA GPU embeddings. Works with Claude, Gemini CLI, Cursor, Windsurf, VS Code Copilot. 30 tools, 1121 tests.

Source