Dragon Brain

English | 中文 | 日本語 | Español | Русский | 한국어 | Português | Deutsch | Français

Memory infrastructure for AI agents — that fails loud, by design.

LongMemEval

License: MIT CI Python 3.12+ Docker MCP Tools Tests Contracts Gauntlet GPU GitHub stars

100% LongMemEval R@5 · 34 MCP tools · sub-200ms hybrid search · CI-gated fail-loud contracts · No LLM required

An open-source MCP server that gives any LLM long-term memory using a knowledge graph + vector search hybrid. Store entities, observations, and relationships — then recall them semantically across sessions. Works with any MCP client: Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Gemini CLI, VS Code Copilot, or any LLM that speaks Model Context Protocol.

Unlike flat chat history or simple RAG, Dragon Brain understands relationships between memories — not just similarity. An autonomous agent ("The Librarian") periodically clusters and synthesizes memories into higher-order concepts.

And it tells you when it can't remember — instead of pretending the memory was never there. (why this matters →)

Quick Start

Prerequisites: Docker and Docker Compose. Detailed setup: See docs/SETUP.md for comprehensive instructions including prerequisites, platform-specific notes, and troubleshooting.

1. Start the Services

docker compose up -d

This spins up 4 containers:

• FalkorDB (knowledge graph) — port 6379
• Qdrant (vector search) — port 6333
• Embedding API (BGE-M3, CPU default) — port 8001
• Dashboard (Streamlit) — port 8501

GPU users: docker compose --profile gpu up -d for NVIDIA CUDA acceleration.

Verify everything is healthy:

docker ps --filter "name=claude-memory"

pip install dragon-brain

2. Connect Your AI Agent

Claude Code (recommended):

claude mcp add dragon-brain -- python -m claude_memory.server

{
  "mcpServers": {
    "dragon-brain": {
      "command": "python",
      "args": ["-m", "claude_memory.server"],
      "env": {
        "FALKORDB_HOST": "localhost",
        "FALKORDB_PORT": "6379",
        "QDRANT_HOST": "localhost",
        "QDRANT_PORT": "6333",
        "EMBEDDING_API_URL": "http://localhost:8001"
      }
    }
  }
}

3. Start Remembering

You: "Remember that I'm building Atlas in Rust and I prefer functional patterns."
AI:  [creates entity "Atlas", adds observations about Rust and functional patterns]

You (next session): "What do you know about my projects?"
AI:  "You're building Atlas in Rust with a functional approach..." [recalled from graph]

What It Does

How It Compares

Feature comparison based on public READMEs as of March 2026. Open an issue if anything is inaccurate.

Benchmark

Dragon Brain scores 100% recall@5 on LongMemEval (ICLR 2025), the industry-standard benchmark for AI memory systems — 500 questions across 6 categories, no LLM required for retrieval.

Note: R@5 (retrieval recall) and QA accuracy are different metrics — shown together for context. Systems marked with QA accuracy use an LLM to generate answers; R@5 measures retrieval only. Dragon Brain's 100% R@5 means the correct evidence sessions appear in the top 5 results for every question.

Per-Category Breakdown

Reproduce It

pip install dragon-brain
docker compose up -d
python -m benchmarks.longmemeval.runner --dataset oracle

Full methodology, raw data, and the journey from 25% to 100%: RESULTS.md

🔥 Use Cases — See It In Action

Same Dragon Brain, different story. Each demo shows a real scenario with example queries and results. Browse all demos →

Every demo uses the exact same Dragon Brain — no plugins, no customization, no domain-specific code. Just data + queries + connections you didn't know existed.

Architecture

graph TB
    Client["Any MCP Client<br/>(Claude, Cursor, Cline, ...)"]
    Server["Dragon Brain MCP Server<br/>34 tools · FastMCP"]
    FalkorDB["FalkorDB<br/>Knowledge Graph · Cypher"]
    Qdrant["Qdrant<br/>Vector Search · HNSW"]
    Embeddings["Embedding Service<br/>BGE-M3 · 1024d"]
    Librarian["The Librarian<br/>Auto-clustering · DBSCAN"]
    Dashboard["Dashboard<br/>Streamlit · Graph Viz"]

    Client <-->|"MCP (stdio/SSE)"| Server
    Server --> FalkorDB
    Server --> Qdrant
    Server --> Embeddings
    Server -.->|"periodic"| Librarian
    Librarian --> FalkorDB
    Dashboard --> FalkorDB
    Dashboard --> Qdrant

• Graph Layer: FalkorDB stores entities, relationships, and observations as a Cypher-queryable knowledge graph. The system is fully async-native, isolating synchronous database drivers in thread pools via AsyncMemoryRepository.
• Vector Layer: Qdrant stores 1024d embeddings for semantic similarity search
• Hybrid Search: Queries hit both layers, merged via Reciprocal Rank Fusion (RRF) with spreading activation enrichment
• Semantic Radar: Discovers missing relationships by comparing vector similarity against graph distance
• The Librarian: Autonomous agent that clusters memories and synthesizes higher-order concepts

Project Structure

Dragon-Brain/
├── src/
│   ├── claude_memory/          # MCP server — 34 tools, services, repositories
│   │   ├── server.py           # FastMCP entry point
│   │   ├── tools.py            # MCP tool definitions
│   │   ├── search.py           # Hybrid search (vector + graph + RRF)
│   │   ├── repository.py       # FalkorDB graph operations
│   │   ├── vector_store.py     # Qdrant vector operations
│   │   ├── librarian.py        # Autonomous clustering agent
│   │   ├── search_advanced.py  # Semantic radar + associative search
│   │   ├── temporal.py         # Time travel queries
│   │   └── ...                 # Schema, embedding, analysis, etc.
│   └── dashboard/              # Streamlit monitoring dashboard
├── tests/
│   ├── unit/                   # 1,027 unit tests (3-evil/1-sad/1-happy per function)
│   ├── gauntlet/               # 139 mutation, fuzz, property-based, concurrency tests
│   └── integration/            # Live-container kill tests via testcontainers
├── docs/                       # Architecture, user manual, runbook, ADRs
│   └── adr/                    # 7 Architecture Decision Records
├── scripts/                    # Docker, backup, health check, e2e tests
│   └── internal/               # 27 migration, verification, and repair scripts
├── docker-compose.yml          # One-command setup (FalkorDB + Qdrant + Embeddings + Dashboard)
└── pyproject.toml              # Python 3.12+, pip install -e ".[dev]"

MCP Tools (Top 10)

All 34 tools are documented in docs/MCP_TOOL_REFERENCE.md.

Forged in Audit

Most open-source memory systems polish the happy path. Here's the bug Dragon Brain shipped to production for two months — and the infrastructure that now exists so it can't come back.

The lie

Before April 2026, the search() pipeline looked roughly like this:

try:
    # ... 6-channel retrieval pipeline ...
except Exception:
    return []

The MCP search_memory tool then transformed [] into the string "No results found." Claude received that string and treated it as authoritative — "the user genuinely has no memories about this topic" — when in reality the embedding service had crashed, FalkorDB was unreachable, or Qdrant timed out.

Every degraded query was the AI operating on missing context without knowing it. A confident lie indistinguishable from genuine emptiness, baked into the system at its most-called function.

The fix

A 4-phase adversarial audit found 83 contract violations across 37 source files. Ten batches of remediation shipped between April and May 2026:

• SearchError is now raised on infrastructure failure — empty list means "no results found", only.
• MCP search_memory returns structured {"error": "MEMORY_LAYER_DEGRADED", "retry_safe": true} — surfaced to the AI as a degradation signal, never a confident lie.
• Cross-store compensation in entity create/update/delete — Qdrant write failure rolls back FalkorDB to prevent split-brain orphans.
• Edge writes use MERGE, not CREATE — retried create_relationship calls don't duplicate edges.
• FTS write failures propagate to caller receipts — silent index staleness eliminated.
• Lock manager raises TimeoutError on contention — never silently proceeds without the lock.
• MCP tools have semantic validation — bad UUIDs return {"error": "ENTITY_NOT_FOUND"}, not silent empty results.

The discipline that keeps it fixed

• tox -e contracts — CI gate baseline-locked at 13 violations (down from 64). New violations fail the build before merge. Quarterly reviews ratchet the baseline toward zero.
• Behavioral integration tests — testcontainers-python spins up real falkordb/falkordb:v4.14.11 and qdrant/qdrant:v1.16.3, then container.kill() mid-operation to assert the fail-loud contract holds end-to-end.
• Async-native repository — AsyncMemoryRepository isolates synchronous DB drivers in thread pools across ~75 call sites.
• Trust-boundary documentation — every cross-process boundary has an explicit contract recorded in docs/ARCHITECTURE.md.

Why it matters

If your memory layer can lie about its failure modes, every downstream reasoning step is corrupt. AI agents trust their tools. Tools that confidently fabricate empty results poison entire reasoning chains.

Dragon Brain is the first open-source memory system we know of with a CI-enforced contract that infrastructure failure cannot be silenced. If it ever happens again, the build breaks before merge.

Receipts

• 1,337 tests across 106 test files, 0 failures, 0 skipped
• Mutation testing — 2,270 mutants, 1,184 killed across 27 source files (3-evil/1-sad/1-happy per function)
• Property-based testing — 38 Hypothesis properties
• Fuzz testing — 30K+ inputs, 0 crashes
• Static analysis — mypy strict mode (0 errors), ruff (0 errors)
• Security audit — Cypher injection audit, credential scanning
• Dead code detection — Vulture (0 findings)
• Dragon Brain Gauntlet — 20-round automated quality audit, A− (95/100)

Full gauntlet results: docs/GAUNTLET_RESULTS.md · Trust boundaries: docs/ARCHITECTURE.md · Integration tests: tests/integration/test_db_kill_scenarios.py

Round 2 (May 2026, v1.2.1) — The Adversarial Auditor

Round 1 installed the contract that prevents silent failure. Round 2 added the auditor that catches the contract being violated.

The AI Council formalized a four-seat trifecta:

• Architect writes specs + audit criteria
• Builder implements per spec
• Auditor verifies per pre-defined criteria — independently of the builder
• Director approves

The auditor doesn't see the build recipe. Auditing recipes biases verification toward checkbox-following instead of outcome-achievement. The auditor sees the bug being fixed and the per-PR criteria; everything else they reconstruct from running the code themselves.

What the new auditor caught that 10 batches of prior Round 1 remediation missed:

• A Cypher label injection vector — unvalidated user input was being interpolated into a graph schema MERGE statement at repository.py:90
• A point_in_time_query silently returning wrong answers — the temporal filter checked a Qdrant payload field that the writer never stored
• A temporal direction enum drift — forward/backward parameter values silently fell through to the default branch because the repository only matched before/after

Three real production bugs, in code that had already passed prior trifecta scrutiny. The auditor seat earned its keep on the first audit pass.

Plus a deterministic-tooling lesson reinforced: the AST contract scanner caught 62 false-positive violations (properly-awaited async calls flagged because the scanner's heuristic didn't check the await keyword) that LLM eyeballing could never have enumerated. The audit protocol now mandates running deterministic tools (tox -e contracts, mypy --strict, bandit, ruff, tox -e integration) before any LLM reasoning. LLM consensus ≠ correctness; deterministic tools are the fourth leg of the table.

Full process artifacts public in process/ as a worked example of the trifecta pattern in production:

• process/REMEDIATION_BUILD_SPEC.md — builder-facing spec
• process/REMEDIATION_AUDIT_SPEC.md — auditor-facing spec
• process/PR_1_HANDOFF.md through process/PR_6_HANDOFF.md — per-PR completion artifacts

Includes the 6-round PR-5 saga where the auditor sent the builder back multiple times — for a test bug masking the actual scenario being tested, a scope-creep refactor that broke dashboard/scripts callers, and persistent handoff hygiene drift — before clean landing.

Updated Receipts (v1.2.1)

• 1,278 unit tests (up from 1,166) — 0 failures, integration tests passing under real-container testcontainers harness
• Contract scanner baseline: 13 (down from 75 — 62 false positives eliminated in PR-6's await-detection fix)
• Cross-store compensation symmetric — both entity and observation create paths roll back graph writes on Qdrant failure
• Channel health observable — search_memory(include_meta=True).metadata.channels exposes per-channel status (healthy/degraded/failed) across all 6 retrieval channels
• point_in_time_query actually works — was producing wrong answers pre-v1.2.1; fixed via payload contract + backfill script
• Cypher injection vector closed — create_memory_type(name=...) validates against a strict regex; defensive assert at the repository layer as belt-and-braces

Why I Built This

Claude is brilliant but forgets everything between conversations. Every new chat starts from scratch — no context, no continuity, no accumulated understanding. I wanted Claude to remember me: my projects, preferences, breakthroughs, and the connections between them. Not a flat chat history dump, but a living knowledge graph that grows richer over time.

Documentation

Local Development

Requires Python 3.12+.

# Install
pip install -e ".[dev]"

# Run tests
tox -e pulse

# Run server locally
python -m claude_memory.server

# Run dashboard
streamlit run src/dashboard/app.py

Claude Code CLI

claude mcp add dragon-brain -- python -m claude_memory.server

For environment variables, create a .env file or export them:

export FALKORDB_HOST=localhost
export FALKORDB_PORT=6379
export QDRANT_HOST=localhost
export EMBEDDING_API_URL=http://localhost:8001

Troubleshooting

ports:
  - "6380:6379"  # Map to a different host port

export PYTHONPATH=/path/to/Dragon-Brain/src

docker exec claude-memory-mcp-graphdb-1 redis-cli GRAPH.QUERY claude_memory \
  "MATCH (n) RETURN count(n)"

More: docs/GOTCHAS.md · docs/RUNBOOK.md

Roadmap

Dragon Brain is under active development. See the CHANGELOG for recent updates.

Current focus areas:

• Native async via falkordb.asyncio.FalkorDB — current asyncio.to_thread wrapper is a correct intermediate state
• Drift detection and quality monitoring for long-lived graphs
• Search result ranking improvements
• Performance optimization for large graphs (10K+ nodes)
• Contract baseline ratchet — quarterly reduction toward zero

Have an idea? Open an issue.

Contributing

See CONTRIBUTING.md for testing policy, code style, and how to submit changes.

Community

• Questions & Discussion: GitHub Discussions
• Bug Reports: GitHub Issues
• Feature Requests: GitHub Issues

License

MIT