Knowledge Graph for Claude Code

Persistent, git-native memory that makes your AI coding agent actually remember. Zero databases, zero services — just bash, jq, and your own commits.

Claude Code and other AI coding agents forget everything between sessions — you end up re-explaining the same project context every time. Knowledge Graph fixes that by turning your file operations and git history into a lightweight, evidence-based memory layer that lives inside your repo.

First-class support for:

• Claude Code — auto-tracks reads and writes via hooks, injects a work snapshot on every session start, rebuilds context after /clear and /compact
• Codex / Cursor / Windsurf / any MCP client — 7 tools and 20+ resources exposed by the bundled MCP stdio server (kg_read_node, kg_query, kg_recent_work, kg_blind_spots, …)

No embeddings. No vector stores. No external services. Works on macOS, Linux, and Windows.

---

Who this is for

• Vibecoders — you describe intent, the agent writes code. Knowledge Graph gives the agent the project context you never had to learn, so one-line requests turn into working changes instead of destructive rewrites. From the maintainer (a vibecoder himself): goal completion and "actually what I wanted" rate jumped at least 10× after installing it — "10× is the floor."
• Senior developers — you want structured, auditable context that your AI agent respects. Every rule traces back to a commit hash or a recorded error event. No hallucinated conventions.
• Teams — rules live in CLAUDE.md right next to the code they govern. Share via git push.

---

Quick Start

macOS / Linux / WSL

bash <(curl -fsSL https://raw.githubusercontent.com/hilyfux/knowledge-graph/main/standalone/install.sh) /path/to/your-project

Windows (PowerShell + Git Bash)

git clone https://github.com/hilyfux/knowledge-graph.git
cd knowledge-graph
.\standalone\install.ps1 C:\path\to\your-project

Then:

1. Restart Claude Code (so hooks activate) — or your MCP-aware agent
2. Run /knowledge-graph init

From that point on: silent tracking, distributed CLAUDE.md knowledge nodes per module, and cross-session memory readable by any MCP-aware agent.

---

vs Alternatives

	Knowledge Graph	mcp-knowledge-graph	Memento	Caveman
Storage	Plain files in your repo	Neo4j database	Vector database	N/A (stateless)
Dependencies	jq only	Neo4j + Node.js + Docker	Python + ChromaDB	Python (optional)
Learns over time	✅ Inference engine	❌	❌	❌
Predicts context	✅ Co-change analysis	❌	❌	❌
Survives clear / compact	✅ Snapshot + @include	N/A	N/A	N/A
LLM cost	Near zero (bash computes)	Every query	Embedding costs	Zero
Team sharing	git push	Manual DB export	Manual DB export	N/A
Multi-agent (Codex / MCP)	✅ 7 tools + resources	Partial	Partial	❌
Windows (PowerShell installer)	✅	❌	❌	❌

---

What You Get

• Cross-agent memory — works natively in Claude Code (hooks); works in Codex / Cursor / Windsurf / any MCP client through the bundled server (7 tools + 22 resources auto-exposed)
• Session-to-session continuity — snapshot survives clear and compact; includes git status uncommitted changes so the agent knows what's still in progress, not just what was committed
• Predict errors before they happen — co-change prediction preloads related-module prohibitions on first access; Read size-guard warns before a 25K-token Read hits its ceiling, so the agent knows to Grep + partial-read instead of burning a round-trip
• Auto-discovered dependencies from real co-change patterns — observe work, infer patterns, promote only evidence-backed rules
• Zero-interrupt workflow — heavy analysis mostly runs at session boundaries; long sessions get a throttled background refresh so graph-analysis.json does not go stale
• Named event channels + schema — parallel streams for domain-specific trackers ({channel}-events.jsonl) with formal event shape and corrupt-line tolerance. See events-schema.md.
• Zero dependencies beyond jq — no Docker, no Neo4j, no Python, no services, no daemon. Inspectable. Versionable. No lock-in.

---

Token Budget

Component	Tokens	When loaded
Knowledge index (pointer tags)	~300-500	Always (@include)
Work snapshot	~200-400	SessionStart / PostCompact
Predicted prohibitions	~100/module	First access to new module
Module CLAUDE.md	~200/module	On file access (lazy)
Total baseline	~500-900	<0.5% of 200K context

---

How It Works (briefly)

Hooks fire silently during your normal Claude Code workflow:

• Read / Write → events recorded in ~3ms; first access to a module triggers a co-change prediction that pre-loads related module prohibitions; long write-heavy sessions also trigger a throttled background refresh of graph-analysis.json
• SessionStart / PostCompact → injects the last work snapshot so the agent picks up where it left off
• Stop → saves the snapshot, rotates the event log, runs background analysis

Pure bash + jq mines patterns from the event log and git history; the LLM is only involved when a CLAUDE.md actually needs to be (re)written. Everything else is zero-token.

Deep dive with full hook table, pipeline diagram, and context-survival matrix: docs/architecture-notes.md.

For non-Claude agents: the same data (CLAUDE.md nodes, work snapshot, co-change pairs) is accessible via the MCP server.

---

Commands

Command	Purpose
/knowledge-graph init	Full project scan. Generates CLAUDE.md for every module.
/knowledge-graph update	Incremental refresh + inference engine.
/knowledge-graph status	Coverage, health, blind spots, activity heatmap.
/knowledge-graph query <question>	Search the graph; get sourced answers.

---

What Gets Generated

Each module directory gets a CLAUDE.md (≤20 lines, maximum information density):

# auth

## Prohibitions
- Raw token in localStorage → XSS (a3f21b)
- Skip refresh in test mock → flaky CI (8c4e01)

## When Changing
- Token flow → @middleware/CLAUDE.md
- User model → @api/users/CLAUDE.md

## Conventions
- Auth errors: 401 + {code, message}
- Refresh tokens: httpOnly cookies only

@ references form the dependency graph. The inference engine discovers and adds them from co-change patterns automatically.

---

MCP Server

7 tools and a resources channel exposed via MCP, usable from any MCP-aware agent (Codex, Cursor, Windsurf, Claude Desktop, custom clients):

Tool	Description
kg_status	Coverage, pending events, blind-spot count, hot zones, recent failures
kg_query	Full-text search across every CLAUDE.md / SKILL.md body — returns path:line:excerpt
kg_read_node	Fetch the full knowledge node for a specific module
kg_recent_work	Current work snapshot — active modules, uncommitted changes, recent commits
kg_predict	Predict related modules for a file path (co-change history)
kg_cochange	Top co-change directory pairs — implicit dependencies
kg_blind_spots	Modules with activity but no knowledge node

Plus Resources: every CLAUDE.md / SKILL.md is exposed at kg://node/<path> (or kg://skill/<path>); the knowledge index at kg://index; the work snapshot at kg://snapshot. Agents that speak the resources/list + resources/read protocol can discover and read knowledge files without knowing filesystem paths.

Auto-registered in .mcp.json during installation.

---

Design Principles

1. Zero interrupts. Never blocks your coding. Analysis runs at session boundaries.
2. Bash computes, LLM decides. Pattern mining is pure bash (~3ms/event); LLM only writes prose.
3. Evidence-based only. Every rule traces back to a commit, error, or analysis. No evidence, no rule.
4. Predict, don't react. Pre-load related knowledge before errors, based on co-change history.
5. Survive everything. clear, compact, long sessions — working state persists through snapshots.
6. Minimal token footprint. ≤20 line CLAUDE.md, pointer-style index, lazy loading.
7. Agent-agnostic outputs. Hooks are Claude Code-specific; outputs (CLAUDE.md nodes, MCP tools, resources) are consumable by any agent.

---

Requirements

• bash — macOS / Linux: native. Windows: Git Bash (winget install Git.Git) or WSL.
• jq — brew install jq / apt install jq / winget install jqlang.jq
• git (optional, recommended) — enhances dependency analysis and evidence tracing
• An MCP-aware AI agent: Claude Code natively, or Codex / Cursor / Windsurf / Claude Desktop via the bundled MCP server

---

Learn More

• Installation — platform-specific setup (macOS / Linux / Windows / WSL)
• Configuration — env vars and tuning
• Architecture — hook flow, prediction engine, pipeline diagram, installed layout
• Events Schema — channel concept + event shape + tolerance guarantees
• FAQ — common questions
• Changelog — release history

---

Contributing

Contributions welcome. See CONTRIBUTING.md.

High-impact areas:

• New pattern types in infer.sh
• Large-monorepo performance (1000+ modules)
• Prediction accuracy measurement and feedback loops
• Integration tests for non-Claude MCP clients
• Additional agent integrations beyond MCP

---

License

MIT