llm-wiki-memory
A local Git‑versioned memory for AI coding assistants that keeps project knowledge persistent so you don’t have to re‑explain every time.
Tool overview
Adoption verdict: llm-wiki-memory is early‑stage (102 GitHub stars) with very few documented real‑world deployments. One detailed case shows it serving as the memory foundation of a personal financial AI analysis pipeline, proving the concept works. The tool is functional but yet to see broader community validation.
What it actually does: It pairs a local Markdown wiki with an MCP server so agents like Claude Code can capture, compile, and recall project‑specific knowledge. Users curate topic pages, and on‑device embedding models (e.g., bge‑m3) enable semantic search over those memories. It solves the “the agent forgets previous context” problem. Think of it not as a cloud RAG service or a LangChain memory backend, but as a local, Git‑friendly Obsidian vault purpose‑built for AI coding agents—knowledge lives as version‑controlled Markdown files the agent can query.
Barriers and costs: The project is open‑source and free to use, with no API charges. You need to set up a Python environment and provide a local embedding model (like Sentence Transformers). CPU‑only inference can be slow; the existing case report uses an Intel iGPU to achieve real‑time retrieval.