Knowledge Graph

Marketplace Active

Persistent memory layer for Claude Code with zero-interrupt architecture

1 Plugin

Purpose

To enable AI coding agents to retain project context across sessions and interactions, leading to more efficient and accurate code generation by providing a persistent, evidence-based memory.

Features

Persistent, git-native memory layer
Zero external dependencies beyond jq
Works natively with Claude Code hooks
MCP server for Codex, Cursor, Windsurf
Session continuity with work snapshots
Co-change prediction for related modules
Local file operations and git history tracking

Use Cases

Maintaining project context for AI coding agents across multiple sessions
Providing structured, auditable context to AI agents to ensure rule adherence
Enabling teams to share AI agent knowledge via shared git repositories
Reducing LLM token costs by performing heavy analysis locally

Non-Goals

Providing a vector database or embedding store
Requiring external services or daemons
Replacing core AI agent functionality with opaque logic
Storing or processing sensitive user secrets

Workflow

User interacts with AI agent (Claude Code / Codex)
Agent reads/writes files, triggering hooks
Hooks record events locally
Background analysis mines git history and events for patterns
Work snapshot saved at session end
Snapshot injected at session start, providing continuity
MCP server exposes knowledge graph data to other agents

Practices

Evidence-based reasoning
Local-first architecture
Developer workflow augmentation

Prerequisites

bash
jq
git (optional, recommended)

Documentation

info:Configuration & parameter referenceWhile the README mentions configuration via env vars and tuning, specific details on all options, defaults, and precedence are not explicitly documented beyond general pointers.

Trust

warning:Issues AttentionThere is 1 open issue and 0 closed issues in the last 90 days, indicating a low rate of issue closure and potentially slow maintainer response.

Code Execution

info:ValidationWhile the bash scripts likely handle arguments, there's no explicit mention or evidence of a schema validation library like Zod or pydantic being used for input sanitization.
info:Error HandlingThe README and architecture notes mention error events and corrupt-line tolerance, but the specifics of error catching, categorization, and structured reporting are not detailed.
info:LoggingThe extension mentions an audit log for destructive invocations and an event log, but details on structured audit records and their accessibility are not fully elaborated.

Errors

info:Actionable error messagesWhile the extension mentions error logging and tolerance, the details on specific error message content (what, why, remediation) are not explicitly provided in the documentation.

Installation

/plugin marketplace add hilyfux/knowledge-graph

Contains 1 extensions

Plugin (1)

Knowledge Graph Plugin

Persistent memory layer for Claude Code — tracks file operations, mines co-change patterns, generates distributed CLAUDE.md knowledge nodes. Zero dependencies beyond jq.