nanograph

Nanograph is an on-device graph database for agents and humans. It stores the graph on the user’s machine and emphasizes a setup described as one CLI, one folder, schema-as-code, and no server.

Its use cases include context graphs for decision traces, causal chains, and structured business context; agentic memory for AI agents; personal knowledge as a semantic layer for bookmarks, highlights, and notes; dependency graphs for packages, services, and lineage tracking; and feature generation from relationships and contextual networks. The product also describes schema-as-code with .pg files that are version-controlled and enforced at query time. It says the graph is a file rather than a service, and that it can be backed up with git or copied with cp. CDC is built in so every mutation is tracked, and the text says this supports rewinding to any point and understanding what changed.

Delivery is centered on a command-line workflow, with installation shown as brew install nanograph/tap/nanograph. The page says a nanograph project is just files in a folder, including schema.pg for schema definition, a .nano folder for graph data using Lance tables, seed.jsonl for rebuilds, and queries for named queries. It also states that the core language is fast, safe, has no runtime, and compiles to a single binary. The query engine compiles to native code and runs against Arrow arrays, and the in-memory format is a columnar data model with zero-copy interop across the data ecosystem. The file storage format is columnar and described as designed for data lake and heavy ML workloads. The text also mentions a full search stack with full-text, semantic, fuzzy, BM25, hybrid, and graph-constrained reranking.

Nanograph is open source. The page presents it as a tool for agents running in CI, in CLIs, on laptops, and even air-gapped, and it says the graph stays private by default because everything remains on the machine. The site also lists GitHub and Read the Docs.