CortexDB
Agent memory usually means a service to run, a vector store to pay for, and facts that go stale without anyone noticing. CortexDB takes the opposite route. It is a pure Go database that puts agent memory and a knowledge graph in one SQLite file, with vectors, RAG, RDF and SPARQL, Cypher, and more than 100 MCP tools, and it works with no embedding model. An optional Claude Code mod writes memories 90 seconds after a session goes idle, and when a new fact contradicts an old one, the old one is kept as superseded instead of being deleted. Version 2.149.0 is the latest release, and the repo now has 480 stars under the MIT license.
What you get:
- Agent memory and a knowledge graph in a single SQLite file, no service to run and no vector store to pay for
- Vectors, hybrid and graph retrieval, plus RDF with SPARQL and read-only Cypher
- More than 100 MCP tools, in-process or over MCP, with Claude Code and Codex plugins
- A 90-second idle capture in the optional Claude Code mod, so session facts become durable memories on their own
- Contradicted facts marked superseded and kept, never silently deleted, with a full history of what changed and why
- Install as a Go library with
go get github.com/liliang-cn/cortexdb/v2, or add the plugin to Claude Code or Codex
Try it on a side project first, and read back what it remembers before you trust it with anything you cannot afford to lose.
Repo: https://github.com/liliang-cn/cortexdb
Quick poll
What would you use CortexDB for first?
- Giving Claude Code a memory that survives between sessions
- Replacing a hosted vector store with one SQLite file
- Querying agent memory with SPARQL or Cypher
- Watching it for now
Vote in the comments.
If you found CortexDB interesting, I write about tools like this every week in The AI-Native Engineering Leader, what actually works for engineering leaders building with AI agents, and what to skip.
Top comments (0)