IWE - Memory system for you and your AI agents

Turn your thinking into queryable context

IWE turns a directory of markdown files into a knowledge graph — a connected structure you browse from your editor and your AI queries from the command line. Same files, same links, two interfaces. No cloud, no database, no lock-in. Version everything with git.

IWE is for people who want database-style queries on their notes — "all drafts under this subtree", "every accepted decision in Q1" — without moving them into an actual database. Write in Markdown, structure with links, give AI agents the tools to navigate your knowledge. IWE itself has no built-in AI — it works alongside Claude, Codex, Gemini, and any tool that speaks the Model Context Protocol.

What You Get

• Plain markdown, full ownership. Your notes are .md files in a local directory. Read them, edit them, git push them. Nothing proprietary.
• A graph, not a folder tree. Link notes together and the same note can belong to multiple topics without copying the file. (How linking works)
• IDE features for your editor. Real LSP integration with VS Code, Neovim, Zed, and Helix — search, refactor, rename, autocomplete.
• Structured access for AI agents. CLI tools and an MCP server give agents parent context and structural navigation over the same notes you edit by hand — retrieval by structure, not similarity guessing.
• Speaks OKF. An Open Knowledge Format bundle is markdown with YAML frontmatter — the format IWE already manages. iwe init --okf scaffolds a conformant bundle, iwe schema validate checks conformance mechanically, and iwe find --filter '{type: …}' queries OKF frontmatter directly.
• Fast. Built in Rust, processes 20,000 files in under a second.

Instructions Aren't Memory

Agents ship with an instructions file — CLAUDE.md, AGENTS.md — and it's tempting to let memory accumulate there too. It doesn't hold. Instructions are rules; memory is what builds up: decisions, corrections, the state of everything in flight. Kept in one flat file, the agent appends, contradictions pile up side by side each looking authoritative, and every session re-reads the whole thing.

IWE stores memory as many small linked documents, so the agent asks instead of re-reading. Most recall is a structured question — what's still open, what did we decide about X, which notes mention this person — answered by a query and one retrieve call that expands the linked context. The rest of the graph stays out of the context window. And because it's markdown in git, the memory is inspectable: open what the agent believes, diff what changed, git-blame when.

How It Works

IWE treats your notes as a connected structure.

... View full README on GitHub