memory-use

English | 中文

Long-term memory for AI coding agents that follows you to every session and every computer. Your notes are plain Markdown in a private git repo you own; this skill teaches Claude Code / Codex to read them before work, write durable facts back after work, keep them synced in the background, and bring everything to a new machine with one line.

Part of the *-use family: profile-use holds who you are, bitwarden-use your secrets, memory-use how your machines and services are set up (NAS, VPN, servers, networks, decisions, rollbacks). Notes store pointers to the other two, never values.

Why not the agent's built-in memory?

	Agent-local memory (~/.claude/projects/*/memory)	memory-use
New computer	Gone	curl … | sh -s -- --repo you/notes
Other agents (Codex, a second Claude session)	Can't see it	Same repo, same skill
You can read / edit / grep it	Hidden files	A normal repo with history
Secrets written by accident	Stay there	Pre-commit hook blocks them

Install

gh auth login   # your notes repo is private

# already have a notes repo
curl -fsSL https://raw.githubusercontent.com/leeguooooo/memory-use/main/install.sh | sh -s -- --repo you/notes

# starting fresh: creates a PRIVATE repo from the starter template
curl -fsSL https://raw.githubusercontent.com/leeguooooo/memory-use/main/install.sh | sh -s -- --repo you/notes --create

This links the skill into ~/.agents/skills, ~/.claude/skills (and ~/.codex/skills), puts memory-use on your PATH (~/.local/bin), clones the notes, enables the secret-scan pre-commit hook, and turns on autosync. memory-use doctor shows the state. Any git host works: --repo git@gitlab.com:you/notes.git.

Also available through the family bundle: /plugin install use-family@leeguooooo-plugins (Claude Code) or install-use-family.sh from leeguooooo/plugins; then run memory-use init --repo you/notes once.

Use

Just talk to the agent: "记一下" / "remember this", "我们之前怎么弄的 VPN" / "how did we set up the VPN", "换电脑了" / "I'm on a new computer". Under the hood:

memory-use brief vpn warp      # context pack: best matching notes + key points, todos, section summaries
memory-use search nas backup    # ranked search; glossary aliases expand each term
memory-use save vpn/servers.md -m "vpn: new exit node"   # secret-scan, commit only these files, pull --rebase, push
memory-use new photos "Photos" "where photos live and how they are backed up"
memory-use lint                # article headers, README length, stale notes, broken links

For a focused topic, brief keeps at most two chunks per note, drops weak matches, and includes only matching todos, summary lines, and commit subjects. A section name or its alias alone (brief nas / brief pan), or no topic, retains the section overview. --max 1500 caps the context body including its truncation notice (minimum 128 characters). Repeating a term or its synonyms does not change its weight; # inside Markdown code fences is not a heading.

Recall hook

Agents forget to look things up. The Claude Code plugin ships two hooks (hooks/hooks.json):

• UserPromptSubmit: when your message names something the notes know (any glossary alias, including colloquial ones like "windows 电脑", or a section name), the agent gets one line: the notes know about leo-desktop — run memory-use brief leo-desktop first. Nothing otherwise; ~50 ms, stdlib only.
• SessionStart: silent unless the notes repo is missing or autosync is stuck.

Installed with install.sh instead of the plugin? Add the same to ~/.claude/settings.json:

"hooks": {
  "UserPromptSubmit": [{"hooks": [{"type": "command", "command": "memory-use hook prompt", "timeout": 5}]}],
  "SessionStart": [{"matcher": "startup", "hooks": [{"type": "command", "command": "memory-use hook session", "timeout": 5}]}]
}

Sync

memory-use autosync on (default after init) runs every 15 minutes via launchd (macOS), a systemd user timer or cron (Linux), or Task Scheduler (Windows):

• fetch, then fast-forward, or rebase local commits onto the remote and push;
• never commits for you and never stashes: an uncommitted edit may be another agent session's half-written note, so it is left alone;
• a conflict aborts cleanly and is reported by doctor and autosync status for a human to resolve.

Agents commit with save, which names its files explicitly, so parallel sessions never sweep up each other's work.

Moving to a new computer

On the old one:

memory-use migrate          # exit 1 if anything is not on GitHub yet: unpushed commits, uncommitted files, stashes
memory-use migrate --push   # push first

It also lists what stays behind by design (the local known-secrets list, the agent's own memory files) and prints the one line to run on the new computer. Keep a setup-new-computer.md in your notes for everything else (SSH keys, VPN, password manager); the template starts one.

Secrets

The pre-commit hook (and save, and check) refuses anything that looks like a secret: private keys, GitHub / OpenAI / Anthropic / AWS / Slack tokens, JWTs, and assignments in English or Chinese (password: …, 密码：…, 路由器密码是 …). Pointers and prose pass (🔑 bw:<uuid>, ~/.ssh/id_ed25519, in secrets.env). With bitwarden-use, memory-use secret put stores a value and prints the pointer, and secret hashes makes the hook catch any password from your vault's memory folder by hash. When profile-use is installed the hook also blocks your personal data.

Configuration

~/.config/memory-use/config.json	{"repo": "you/notes", "dir": "~/github.com/notes"}, written by init
MEMORY_USE_DIR / MEMORY_USE_REPO	override the checkout path / repo
~/.config/memory-use/known-secrets	literal secrets to catch, one per line, never committed
MEMORY_USE_NO_UPDATE_CHECK=1	silence the once-a-day "new version" line

Stdlib Python 3.9+, no dependencies. python3 -m unittest discover -s tests. Release: scripts/release.sh <version> [notes] — bumps, tests, tags, publishes, and syncs the plugin marketplace.

License

MIT