CortexDB

The long-term memory layer OpenHuman’s Memory v2 runs on — and a naming minefield on GitHub.

What it is (the .ai one, the one that matters here)

CortexDB (cortexdb.ai) — “The Long-Term Memory Layer for AI Agents,” the engine behind OpenHuman’s Memory v2 (Recall/Fetch/Store; the hosted CortexDB is TinyHumans’ default, self-hosted CortexDB the alternative). Storage engine built by Prashant Malik, co-creator of Apache Cassandra.

• Memory model: one recall returns events, episodes, facts, beliefs and understanding in a single round trip — HNSW + BM25 fused by RRF; context comes back with valid time and source, “not just similar text.”
• Deployment: Raft-replicated RocksDB, on-prem or their cloud; hierarchical scopes org→agent, PASETO capability tokens, hash-chained audit; reached from Python, TypeScript, CLI, REST, or MCP.
• Self-hosting is the recommended mode (docs): “run the same server the cloud runs, on your own infrastructure — your data never leaves your network, and there’s no per-token cost.” One docker run, ~80 MB Linux binary (amd64+arm64), port 3141, SHA-256 on every release, dashboard at localhost:3141. Fails closed without an embedding credential; loopback-bind + bearer-key hardening in the README.

Is it open source? No — source-available-adjacent, actually just closed-source-with-free-binaries

The site’s own words on the download page: “Source-available, not open source” — and the GitHub org’s README is blunter: “The CortexDB source code is closed. Only prebuilt artifacts live here.”

• cortexdbai/cortexdb-releases — the official repo: LICENSE.txt (Apache-2.0) + README + release binaries. 0★, no source. v0.10.4 (Oct 3, 2026).
• So: free to self-host (binary/Docker, no account, no email), but not inspectable, not forkable. License = Apache-2.0 for the repo’s artifacts/readme, not an OSI grant over unseen source.

The name collision (the GitHub warning was right)

Searching “cortexdb” on GitHub returns many unrelated projects:

• liliang-cn/cortexdb (270★, MIT, Go) — different project entirely: “AI memory and a knowledge graph in one SQLite file. Pure Go: vectors, RAG, agent memory, RDF/SPARQL, Cypher, 80+ MCP tools. Works without an embedding model.” Actively updated (v2.119.0+), legit on its own merits — and ironically more open (real MIT source) than the .ai product. Deep-dive (captured, README in full):

• One SQLite file, no service — pure Go library; postgres:// DSN swaps in PostgreSQL+pgvector. Default brain at ~/.cortexdb/cortexdb.db.

• Works without embeddings: CORTEXDB_EMBED_BASE_URL unset = lexical mode (FTS5 + graph) — a homelab win, runs fully local against llama-swap’s OpenAI-compatible endpoint or no model at all.

• The stack inside: HNSW/IVF/flat/binary vectors + FTS5 hybrid + graph retrieval; RDF 1.2 knowledge graph with SPARQL 1.1/1.2, RDFS + OWL 2 RL inference, SHACL validation — passing the W3C test suites — plus read-only Cypher and a Palantir-style ontology with governed actions; 80+ tools in-process or over MCP; serve_graph_3d live 3D graph view; change feed with exactly-once ordered writes.

• Distribution is the interesting part: Claude Code plugin (/plugin marketplace add liliang-cn/cortexdb) with /remember, /recall and an auto-recall hook; Codex plugin; a cortexdb-live Claude Code mod — planned recall (model turns the prompt into keywords/aliases/entities before searching; a bare “ok” searches nothing), status line (which brain, node count, last recall), and idle-capture (90s after idle, the model distils the session into durable memories under stable slugs — later passes replace, not duplicate). Point several agents/machines at one cortexdb-grpc (default 127.0.0.1:47821, bearer token) for a shared brain.

• Homelab test candidate: single static binary, SQLite persistence, loopback gRPC, no forced cloud — fits the doombox lane (see Local archives & crawlers SQLite-as-agent-exchange thesis).

• AubakirovArman/CortexDB (4★, Apache-2.0) — “agent-native context database prototype”, unrelated.

• ~7 more 0–4★ same-name repos: unrelated prototypes.

Vault kin

• openhuman (OSS) — Memory v2’s engine; the “bring your own CortexDB (endpoint+key)” option in Connections > Memory.
• Agent knowledge systems — the hosted/self-hosted memory-engine lane: Cognition’s Agent Memory Repo (git+markdown) vs CortexDB (Raft-RocksDB service) = the two ends of the memory-substrate spectrum (text-first vs database-first).
• Agent-accessible product — MCP as a first-class client surface.
• Small-language-models lane: works without embedding models (the Go namesake’s trick) vs CortexDB’s OpenAI-embeddings-by-default.

Practice (homelab lens)

• Want to try: docker run on doombox, port 3141, loopback-bind + CORTEX_API_KEY; needs an embedding credential (OpenAI by default — or point at a local OpenAI-compatible server).
• Closed-source caveat for the fleet: free self-host ≠ auditable; keep the Local archives & crawlers SQLite-first lane as the fallback substrate.