Jeviphy: selected claude-mem observations in a code graph (offline pilot)

[luisroquette]

I built Jeviphy, a Graphify-based skill for Claude Code and Codex that links selected project memories to code-graph results. It reads observations through a local project-scoped memory gateway, leaves claude-mem storage alone, and keeps optional Jev relevance calls off by default.

The one-command demo uses synthetic code and memory. It shows a cited auth function moving from rank 3 to rank 1 after a hand-written decision, with zero Jev calls. That is wiring proof only. I have 30 draft questions from a private project awaiting independent gold labels and a review of what may leave the machine. There is no measured real-memory gain yet.

I'd value feedback on the integration boundary: is there a supported project-scoped way to fetch a small, explicit set of observations with stable IDs and provenance? The current gateway stays outside claude-mem. I'm happy to adapt the bridge if there is a better public interface.

The Graphify upstream pilot is a separate draft PR.

[luisroquette]

Follow-up: Jeviphy now has an autonomous 30-question source-provenance audit: https://github.com/luisroquette-labs/jeviphy/blob/jeviphy/docs/jev-memory-pilot-evaluation.md . Of 26 questions linked to an implementation file present in the pinned graph, only six Graphify top-ten candidate pools contained a node from a file cited by the selected memory. This is an exact file-path proxy, not a relevance or Jev quality score. No real memory was sent to Jev. It suggests we should first improve candidate coverage. I am still interested in a supported, project-scoped export with stable observation IDs and source provenance.