hilyfux

Knowledge Graph — Git skill for Claude Code

Git community intermediate

Built on Anthropic internal engineering practices and Karpathy's AutoResearch methodology.

How to install Knowledge Graph

This entry records only its repository, not the path inside it, so there is no exact command to give. Open hilyfux/knowledge-graph and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Knowledge Graph does

**Persistent memory that makes Claude Code smarter with every session.**

Alternatives in Git

Git Worktree Docs Issue: Dependency Installation — GitHub — Community feedback on official docs with full worktree lifecycle walkthrough 81.1k ★
Anthropic 课程 — by anthropics - 官方课程材料 19.7k ★
Claude Relay Service — GitHub Repo stars Self-built Claude code mirror service supporting multi-account switching Relay service 9.9k ★

README

Knowledge Graph for Claude Code

**Persistent memory that makes Claude Code smarter with every session.**

[](https://github.com/hilyfux/knowledge-graph) [](LICENSE) [](https://github.com/hilyfux/knowledge-graph/commits/main)

Claude Code forgets everything between sessions. Knowledge Graph fixes that. It silently tracks every file operation via hooks, builds distributed `CLAUDE.md` knowledge nodes across your project, and injects the right context when Claude needs it -- automatically.

**Zero dependencies beyond `jq`. No databases. No vector stores. No external services. Just bash scripts and git.**

What's New in v1.2.0

Zero-Interrupt Architecture

Previous versions interrupted your coding every 15 writes to force a knowledge update. v1.2 eliminates all mid-coding interrupts. The system now works exclusively at session boundaries:

During coding: only records events and predicts on first module access (~5ms). No blocking, no forced updates, no popups.
On session end (Stop hook): saves a structured work snapshot — which modules you touched, what you modified, errors you hit, commits you made.
On next session / clear (SessionStart): injects the snapshot so Claude immediately knows what you were doing, even after clear.
On compact (PreCompact): saves snapshot before compaction, rebuilds context from snapshot + working set afterward.

Session Recovery That Actually Works

Scenario	Before v1.2	After v1.2
After `clear`	Claude only knows project rules. Forgets what you were doing.	Claude knows which modules you touched, what you modified, errors you hit, and recent commits.
After `compact`	Loses working state. Generic event summaries re-injected.	Working set prohibitions pinned. Snapshot restored with specific modules and error context.