Claude Note CLI
Claude Code and Codex hooks search your notes and surface a known answer before a reply and after a failed command.
- 88 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Reads your session transcripts and saves digests as notes on disk
- Changes hook settings and the `CLAUDE.md` and `AGENTS.md` files and adds an hourly job
Install
- Works in
Manual install
git clone https://github.com/crimeacs/claude-note.git ~/src/claude-noteThen `scripts/install-from-checkout.sh --non-interactive --vault ~/Documents/notes --author you@example.com --claude-hooks` and `claude-note setup --codex`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Reads your session transcripts and saves digests as notes on disk
- Changes hook settings and the `CLAUDE.md` and `AGENTS.md` files and adds an hourly job
This is third-party code. Review the repository files before installing.
What it does
Claude Note captures what Claude Code, Codex, Cursor and ChatGPT sessions learn, keeps it as typed Markdown notes and puts the right notes in front of the agent before it acts: on every prompt and the moment a command fails. Hooks queue every session, and a background worker extracts decisions, gotchas and patterns into notes. Four working modes: recall before an answer (a hook searches notes and adds a short source-linked shortlist, a miss is announced), a sentinel (when a command fails it surfaces the known fix and flags auth, quota and disk-full errors as do not retry), `claude-note remember` for writing a lesson by hand and `claude-note doctor` to check setup. `claude-note eval` measures whether the right note reaches the answer.
Who it is for. Teams and solo developers whose agents step on the same rakes again and forget the outcome of past sessions.
Good fit when
- The agent is about to run a build on a server where it breaks the app, and a team note forbids it.
- A command fails with an error whose fix is already recorded, and it must appear before a retry.
- A lesson from debugging must be written as a `gotcha`, `pattern` or `decision` with one command.
- You need to measure how often recall picks the right note for questions drawn from your own notes.
Not a fit when
- Shared team memory in the cloud is needed: Claude Note keeps notes in your own on-disk vault.
- Session text must not go to external models: synthesis needs the Claude CLI and an explicit source policy.
- A search index and a schedule are unwelcome: installation adds hooks, rules in `CLAUDE.md` and an hourly refresh.
Example request
Search the notes for what we already know about a 403 after adding a Gmail scope and show the source.Limitations
Needs Python 3.11 or newer, `uv`, QMD for search and the Claude CLI for synthesis. Install from a checkout with `scripts/install-from-checkout.sh`, then `claude-note setup --codex`. Setup is idempotent and backs up every file it changes, and `--dry-run` shows the plan. It adds hooks, a rule to search notes before acting in `~/.claude/CLAUDE.md` and `~/.codex/AGENTS.md` and an hourly index refresh. External model calls need an explicit source policy, and existing notes count as unverified leads. Codex hooks run after you trust them on first launch.
How to disable. Remove the hooks and rules added by `claude-note setup` from Claude Code and Codex settings and from `~/.claude/CLAUDE.md` and `~/.codex/AGENTS.md`. Cancel the hourly refresh; the notes vault is deleted by hand.
FAQ
How do I check that it works?
`claude-note doctor` separately checks configuration, isolated hooks, the actual server and recorded delivery. Without it a configured setup is easily mistaken for a working one.
How do I write a lesson by hand?
With `claude-note remember --type gotcha --tags oauth "text"`. Types: `gotcha` for failures, `pattern` for a method and `decision` for an accepted choice.
What does the sentinel do on a failure?
It shows the known fix before a retry. Auth, quota and disk-full errors are flagged as do not retry.
What happens to an existing note on a repeat write?
If a note with the same title exists, `remember` appends a dated block and does not overwrite. Recall shows the newest block first.
How do I measure the benefit?
With `claude-note eval probes -n 80` and `claude-note eval run`: the first writes questions from your notes, the second checks whether recall injected the right note.
Discussion
No questions or reviews yet.
Related
CLI toolsMemory for Claude Code: hooks record human and agent decisions, and a grade is assigned afterwards once the outcome is known.
hipocampus gives an agent file-based memory: a 3,000-token topic index loads each session and shows what came up before
A Claude Code workspace folder: hooks with a local Ollama detect the project type, pick skills and remind you to update the docs
A CLI that indexes memory, rules and sessions of Claude Code, Codex, Cursor and Gemini and hands durable decisions to every agent
agentcairn keeps Claude Code, Codex and Cursor memory in Markdown files for Obsidian and hands an agent only what it needs
Persistent memory for coding agents: insights are stored as markdown cards and recalled at the start of the next session.
README in short
Covers installation, a version 2.1 overview, the capture, recall and sentinel modes, everyday commands, recall quality measurement and the upgrade policy.