# LLM Wiki Memory

A local, git-versioned memory system for AI coding agents that captures conversations, compiles them into durable project knowledge, and recalls context through a local MCP server, stored as plain Markdown with on-device embeddings, no database.

LLM Wiki Memory gives an AI coding agent persistent memory across sessions with no external service: it captures conversations, compiles them into durable project knowledge and lessons the agent applies next time, and recalls the right context through a local Model Context Protocol server. Memory is stored on disk as plain Markdown in a git-versioned wiki, searched with local embeddings rather than a vector database.

An hourly scheduled job, run through cron or launchd, gradually warms embeddings for new notes in the background; without it, recall still works but a newly written note is embedded the first time a search touches it. Bootstrap scripts for macOS, Linux and Windows set the project up idempotently, so re-running preserves existing configuration and rule files, and can optionally git-track the wiki so it is committed alongside the code.

Different moments in the memory lifecycle are automatic, agent-led as part of its normal flow, ask-first before saving, or run entirely in the background, so the user stays in control of what actually gets saved while everything remains on the local machine.

Website: https://github.com/ctxr-dev/llm-wiki-memory