graphify

103,187 stars and 10,024 forks by 6 August 2026, on a repository created 3 April 2026: graphify's adoption curve is the first thing anyone notices. «Safi Shamsi» built it, «Graphify Labs» publishes it, and the thing itself is a skill and CLI that turns a mixed project (code, documents, PDF, images, video, «MCP» configurations, package manifests) into a knowledge graph you traverse instead of grepping files. It is Apache-2.0, shipped on PyPI as `graphifyy`, with a README in 33 languages.

Three design choices carry the argument. Code is parsed locally into «tree-sitter» ASTs with no model in the loop, so a pure code corpus needs no API key. Every edge is labelled by provenance, extracted or inferred. And the project refuses embeddings and a vector store outright. «NetworkX» and the Leiden algorithm do the assembly.

The benchmarks reward careful reading. On LOCOMO, graphify leads recall@10 at 0.497 against 0.149 for supermemory and 0.048 for mem0, yet trails on QA accuracy: 45.3% versus supermemory's 49.7%. On LongMemEval-S it matches a dense RAG at 76%. The defensible edge is cost and traceability, not answer quality.

The presentation around it is less exact. «graphify.net», Shamsi's own site, advertised 3.7k+ stars the same day the API returned 103,187, declared MIT three times against the repository's Apache-2.0 LICENSE file, and still promoted a token-reduction figure that belongs to the v1 README. The default branch is v8; the one on main describes an earlier product entirely.