graphify

«Safi Shamsi» built graphify as a skill and CLI that turns a whole project into a traversable knowledge graph, invoked as `/graphify` from «Claude Code», Cursor, Codex, Gemini CLI and around fifteen other clients. Code is parsed locally into «tree-sitter» ASTs with no model in the loop, and every edge carries its provenance: `EXTRACTED` when it is explicit in the source, `INFERRED` when graphify resolved it. Published by «Graphify Labs» under Apache-2.0, the repository was created on 3 April 2026 and showed 103,187 stars and 10,024 forks on 6 August 2026.

The design refuses embeddings and a vector store. Clustering runs through «NetworkX» and the Leiden algorithm; coverage reaches SQL, Terraform, Apex, «MCP» configurations, package manifests, Office files, PDFs, images and locally transcribed video. `# WHY:` comments become nodes wired to the code they explain.

The benchmark picture is split. On LOCOMO, graphify leads recall@10 at 0.497 against 0.149 for supermemory and 0.048 for mem0, yet trails on QA accuracy at 45.3% versus 49.7%. On LongMemEval-S it matches a dense RAG at 76%. Graph build costs zero LLM credits, which locates the defensible advantage in cost and traceability rather than answer quality.

Then there is the drift. On the same day, «graphify.net», Shamsi's own showcase and directory, advertised "3.7k+ GitHub stars", declared MIT three times against the repository's Apache 2.0 `LICENSE`, and still pushed a "71.5× token reduction" figure retired with the v1 README. A tool vendor writing the comparison guide for its own category occupies the evaluation query first.