RNR Enhanced Cognee¶
Enterprise-grade AI memory infrastructure with multi-tenant support, 30 pluggable storage providers across 4 tiers, 122 MCP tools, and 4 cross-language client SDKs.
What is RNR Enhanced Cognee¶
RNR Enhanced Cognee is a fork of Cognee that adds production-grade infrastructure: multi-tenant data partitioning, 30 pluggable storage providers, four official client SDKs, MCP server hardening (API-key + rate limit + payload cap), HTTPS/TLS support, encryption at rest, GDPR tooling, backup + recovery, and SigNoz + Apache Superset observability.
It exposes 122 MCP tools (Model Context Protocol) for AI IDEs like Claude Code, Cursor, and Continue, plus a FastAPI HTTP variant for cross-language SDK use.
Default stack -- 100% MIT + Apache-2.0¶
| Tier | Default provider | License |
|---|---|---|
| Graph | ArcadeDB | Apache-2.0 |
| Vector | Qdrant | Apache-2.0 |
| Cache | Valkey | Apache-2.0 |
| Relational | PostgreSQL + pgvector | PostgreSQL License |
Every tier accepts a drop-in alternate via ENHANCED_*_PROVIDER env vars -- see Pluggable DB Backends for the full matrix.
Quick links¶
- Quick Start -- install in 30 seconds + first MCP call
- Pluggable DB Backends -- 4-tier factory architecture + how to add adapters
- Profiles -- 5 pre-baked deployment profiles + per-adapter caveat tables
- Deployment Quickstart -- local + VPS deployment runbooks
- Monitoring -- SigNoz + Apache Superset observability stack
- Secrets Management -- four levels of secrecy from laptop dev to compliance-driven
- SDK Publishing -- how to push Node / Rust / Go SDKs to npm / crates.io / pkg.go.dev
- Feature -> License Matrix -- per-feature licensing verdict for downstream commercial use (MAS etc.)
- License Audit -- per-component license analysis
- Commercialisation Guide -- compliance considerations for paid use
Key features¶
Memory + Knowledge Graph¶
- 122 MCP tools across 25+ categories (CRUD, search, dedup, summarisation, GDPR, backup, undo/redo, ...)
- Semantic search with vector embeddings (Qdrant + pgvector + LanceDB + Chroma + Weaviate + Milvus)
- Knowledge graph ingestion (
cognify) with 9 graph backends including ArcadeDB / Apache AGE / Memgraph / Kuzu - Memory deduplication, summarisation, consolidation, tiering, provenance, confidence scoring, importance scoring
- Versioning with undo / redo
- Real Redis-backed personalization (agent affinity + interaction recency + query affinity)
Production Infrastructure¶
- Multi-tenant data partitioning (
TenantContext+ per-tenant Postgres schema + HTTP X-Tenant-ID middleware) - Cross-language client SDKs: Python on PyPI; Node / Go / Rust registry-ready
- MCP server hardening: API-key auth + per-tool token-bucket rate limit + payload cap
- HTTPS/TLS support on the FastAPI server
- Encryption at rest (Fernet AES-128-CBC) with key rotation
- Audit logging + GDPR (consent, export, delete, tenant-isolation verifier)
- Backup + verification + recovery + transaction manager
- 4,661 tests passing, 95% coverage
Observability¶
- SigNoz APM (replaces Grafana / Loki / Tempo / Jaeger -- all Apache-2.0)
- Apache Superset BI with 5 importable dashboards
- Prometheus metrics, OpenTelemetry tracing
- LLM cost tracker with per-model budgets
Performance¶
Real Locust benchmark against the live default stack (ArcadeDB + Qdrant + Valkey + PostgreSQL):
| Metric | Value |
|---|---|
| RPS | 48.86 |
| p50 latency | 11 ms |
| p95 latency | 14 ms |
| p99 latency | 17 ms |
| Error rate | 0.00% |
60s run with 20 concurrent users + 5/sec spawn rate. See tests/benchmarks/baselines/.
Get started¶
git clone https://github.com/vincentspereira/RNR-Enhanced-Cognee.git
cd RNR-Enhanced-Cognee
docker compose -p RNR-Enhanced-Cognee -f config/docker/docker-compose-enhanced-cognee.yml up -d
python -m uvicorn src.enhanced_cognee_mcp:app --port 8080
Then point Claude Code's MCP config at bin/enhanced_cognee_mcp_server.py. Full setup in Quick Start.