LightRAG setup for Claude Code and OpenClaw memory

LightRAG is presented as a shared knowledge base for Claude Code, OpenClaw and other agents that support MCP. The speaker compares it with built-in memory, Obsidian plus skills, Mem0 and GraphRAG, arguing that separate memory files lose connections between facts and consume too much context. These are his assessments and test results, rather than universal findings about those tools.

The demonstration uses project notes and video transcripts to build a graph of entities and relationships. Queries return project details with document references. A Telegram voice note enters the same knowledge base through OpenClaw, and Claude Code later retrieves the client request. The speaker also shows agents saving decisions through configured memory instructions.

The setup deploys the open source software through Docker on a server and connects agents through MCP. The speaker reports roughly ten minutes for his instruction-driven installation, though he notes that server configurations can require fixes. His self-hosted deployment uses an external model API for extraction and answers, plus an embedding model; it does not demonstrate fully offline inference. The installation instructions and agent configs are offered through his private Telegram channel.

Import quality matters: raw logs produced unwanted entities in his graph. He recommends starting with ten useful facts, recording reasons for decisions, and cleaning voice notes before saving them.