Privacy-first & self-hosted
NovaCortex runs entirely on your own infrastructure with a single Docker deployment. Your knowledge base never leaves it — no third-party cloud, no data egress you did not approve. Retrieval even sends less data per query to the model provider than dumping raw context.
Portable open format (PMF)
Export and import the entire graph as human-readable JSON, binary MessagePack (~60% smaller), or AES-256-GCM encrypted bundles. Every export ships with Merkle and content-hash integrity. Take your memory anywhere — zero lock-in.
Graph-native memory
Typed memories — episodic, semantic, procedural, and working — linked by a real relation graph. Edges like causes, contradicts, supersedes, and same_as model how knowledge actually connects, not just a flat pile of embeddings.
Semantic + text retrieval
OpenAI-embedding vector search, or any OpenAI-compatible / local endpoint you point it at, with a transparent text fallback when embeddings are unavailable. Add optional recency weighting and rely on strict namespace isolation between projects.
Plugs into your agents
A native MCP server exposes 15 tools to Claude, Cursor, and any MCP client. Prefer code? Reach for the REST API (with Swagger), TypeScript and Python SDKs, or the CLI. Wire NovaCortex into whatever stack you already run.
Context efficiency
Retrieval keeps prompts lean — up to 98.8% fewer input tokens than dumping the whole knowledge base into context, and 80.0% on LongMemEval_S at ~5% of the full-context token bill. It is the proof, not the pitch: smaller prompts mean lower cost and sharper answers.
See the benchmark →Ops-ready
Opt-in OpenTelemetry traces, scope-gated API tokens, an audit log, and webhooks for everything that matters. Background processing handles decay, consolidation, and relation discovery so the graph stays healthy on its own.