Local-first memory for
Claude Code.

See it remember

A fresh session, a vague request — and your assistant already knows your project's conventions. No re-explaining.

What makes it different

Plenty of tools store notes. REKOL is the one that bolts onto the assistant you already use, needs no key, and brings the right memory to you.

And the basics, done right: memory is plain markdown you own (open it in Obsidian, grep it, version it in git), and REKOL searches your past session transcripts alongside your curated notes. It's also time-aware — REKOL keeps the assistant grounded in the actual current time every turn, not just the date its session started with, and weights recall by recency, so the most current memory surfaces first.

Install

Works with Claude Code — terminal, IDE, or the Claude Desktop app — on macOS and Linux today. It's git clone of a public repo you can read first — not a script piped into your shell.

# clone and install — it asks where to keep your memory
git clone https://github.com/rekol-io/rekol && cd rekol && ./install.sh

install.sh prompts for your memory folder (default ~/rekol-memory — press Enter to accept), then sets up a local venv + the rekol CLI, wires the Claude Code hooks & skill, builds the index, and makes your existing Claude Code history searchable right away. The curated starter set — the always / when / topics memory — comes from the assistant-led “set up my rekol memory” step after install, not the installer. Sync that folder via Dropbox / iCloud / git / Syncthing, or keep it local.

Bring in your history — you stay in control. Your past sessions become searchable at install (locally — your transcripts are never uploaded). The curated layer is opt-in and review-gated: just tell your assistant “set up my rekol memory” (or run rekol init yourself), and you approve everything before it's kept.

Changed your mind? ./uninstall.sh reverses everything cleanly (and backs up what it touches) — your markdown memory always stays yours.

How it works

Your memory is markdown in folders you own. It ships with sensible default layers — REKOL indexes everything locally and surfaces the right pieces to your assistant.

Smart defaults, not a cage — it's your folders, so add your own or reshape them to fit how you think.

You don't run memory commands — you just work, and the assistant already has your context and uses it.

You:    Add rate limiting to the login endpoint.
Claude: Done — RS256 auth, your Problem+JSON error format, and the
        retry rules you settled on last week. (recalled from
        memory automatically — you never re-explained any of it)

REKOL loads relevant memory at the start of every session, and the assistant pulls in more from your notes and past sessions on its own as it works. It feels like the assistant simply knows your project and thinks with it. You can steer it when you want — "remember this", "what did we decide about X?" — but you rarely need to.

Your memory lives on your machine

Not by promise — by architecture. REKOL stores and searches it locally, with no server, account, or telemetry of its own.

The honest picture: REKOL stores, embeds, and searches your memory entirely on-device — it phones home to no one. The only time REKOL touches the network is a one-time embedding-model download (BAAI bge-small, ~130 MB from Hugging Face) on first run. Your AI assistant is, of course, still its own cloud service — when it uses your memory, that context goes to its provider just like anything else you type. REKOL doesn't change that; it just adds no new cloud, key, or telemetry of its own.