CortexDB

CortexDB

A pure-Go, single-file AI memory and knowledge graph. One SQLite file holds vectors, hybrid RAG search, scoped agent memory, an RDF/SPARQL knowledge graph, a Palantir-style ontology, and 60+ agent tools — embedded in your Go program, or installed as a shared brain for Claude Code / Codex. Works with no embedder (lexical mode, no API key) or any OpenAI-compatible embeddings endpoint. No service to run.

go get github.com/liliang-cn/cortexdb/v2

_{serve_graph_3d on a real shared brain — the one behind an OpenClaw cluster: 2000 entities, 5953 relations, node types the agents wrote themselves. Served from inside the MCP server handling the calls, so the graph lights up as tools touch it. (Orbit, as MP4)}

db, _ := cortexdb.Open(cortexdb.DefaultConfig("brain.db"))
defer db.Close()
brain := db.KnowledgeMemory()
_, _ = brain.Remember(ctx, cortexdb.KnowledgeMemoryRememberRequest{Content: "Alice prefers tabs.", Scope: "user"})
rec, _ := brain.Recall(ctx, cortexdb.KnowledgeMemoryRecallRequest{Query: "what does Alice prefer?"})
fmt.Println(rec.ContextPack.Text) // paste-ready context pack with source attribution

What's inside

• KnowledgeMemory brain facade — Recall / Remember / Reflect / Consolidate / PromoteToKnowledge / context packs; fused retrieval across episodic memory, durable knowledge, and GraphRAG chunks; relational answers returned as graph facts (Alice —uses→ Apollo) read from edges, reliable even with no embedder; deterministic no-LLM extract_conversation; memories can carry inline entities/relations so one call stores and graphs them.
• Composable retrieval — cortex_query: vector / lexical / hybrid / graph prefetch lanes fused by RRF, weighted RRF, or DBSF, with metadata filters and per-source score debugging; an Authorize callback gates every candidate (RBAC/ABAC at the retrieval layer); pluggable reranker.
• Vector + lexical engine — FTS5, HNSW / IVF / Flat indexes, scalar & binary quantization, geospatial indexing, semantic query routing.
• External retrieval lanes — a search cluster you already run (Meilisearch, Weaviate, …) can be one fused lane via QuerySource, without becoming the storage: it names candidate ids, the brain still owns the content, and a stale id is dropped rather than fabricated.
• The knowledge contract — every record can say how it knows and how sure: _source, _chunk, _producer, and a _grade from a closed set — verified (a named person kept it), self_consistent (derived from something that stated it), asserted (a model or a person said so, unchecked), held, refused (the vocabulary declined it, with why). contract_tally answers what the whole shelf stands on, counting the untagged rows too; contract_needs_attention lists what wants a person; fact_provenance cites the text a fact came from. Producers call ValidateContract before they write. alchemy writes the contract for every graph it loads here — the one sink of its six that does.
• Swappable storage — SQLite by default; a postgres:// DSN moves the same brain to PostgreSQL + pgvector, with vectors, hybrid search, memory and the RDF graph all running on either. Compile-time backend registry, not a plugin system (storage is the hot path). 104 opt-in PostgreSQL tests, mostly parity: one test body, both databases, same answer required.
• Knowledge graph — RDF triples/quads on the same file: a practical SPARQL subset (updates, OPTIONAL/UNION/VALUES, aggregates, subqueries, property paths), RDFS-lite materialized inference, SHACL-lite validation, N-Triples/Turtle/TriG I/O; property-graph apply_inference materializes two-hop relation compositions with provenance; entities track asserting documents, and delete_document_graph is deletion shaped like ingest.
• Ontology (Palantir-style) — typed object/link/interface types with primary keys and cardinality, an object-set algebra (union / intersect / filter / search_around), governed action types with audit trail, generated typed agent tools, and a breaking-change schema diff; strict or vocabulary enforcement.
• Pipelines — memoryflow (transcript → recall → wake-up → promotion), graphflow (corpus → graph → HTML report), importflow (CSV / SQL dumps / live Postgres-MySQL → RAG + KG), connector (PII masking, signed plans, reversible vault, CDC sync).
• Tools & MCP — 60+ tools with the same names in-process and over MCP, plus render_graph_html, an interactive graph view.
• Quality, measured — pkg/eval runs a labeled query set through the real retrieval path with recall@k / nDCG regression floors in CI; FTS5 / SPARQL / SQL-dump parsers are fuzz-tested.