Agent Carnet

Agent Carnet

A tiny CLI that gives you a shared markdown notebook on disk under .carnet//.md. Notes have a 30-day default lifespan that resets every time they are read or applied; useful ones survive, stale ones drift to .trash/ automatically.

Quick reference

# Save (always pass --summary and --agent claude-code)
echo "body content" | agent-carnet save deps/iconv-issue \
  --summary "iconv-esm v0.7 types broken — pin to v0.6" \
  --agent claude-code \
  --tags compat,esm

# Recall
agent-carnet find iconv               # search summaries (does NOT bump lifespan)
agent-carnet list                     # category-grouped overview, sorted by last_used
agent-carnet list --sort use_count    # most-applied notes first
agent-carnet read deps/iconv-issue    # read full content (bumps last_used; weak use signal)

# Mark as actually applied (strong use signal — bumps last_used + use_count)
agent-carnet used deps/iconv-issue

# Maintain
agent-carnet move  
agent-carnet rm  --yes

save reads body content from stdin (pipe via echo, heredoc, or -h (e.g. agent-carnet save -h, agent-carnet used -h). Each subcommand prints its own focused help — required arguments, options, and examples — without invoking filesystem operations.

When to save

Treat carnet as an agent notebook, not a strict archive. If you think a note may help a future agent (or future-you) resume work, dodge a rethink, remember a temporary state, or preserve a useful hunch — save it. Provisional notes are welcome; the 30-day lifespan is the safety valve (notes that keep getting read or applied stay alive, the rest expire automatically).

Examples worth keeping in mind (not an exhaustive list):

• "We tried X. It didn't work because Y. Don't waste time retrying without Z changing."
• "The library does A, but only if you also call B beforehand. Not in their docs."
• "Chose approach P over Q. Q would have been faster but breaks our R constraint."

These are guidelines, not gates. If grep / read / git log already gets a future reader to the same conclusion, you can skip it — but when in doubt, prefer saving a small, well-summarized note over losing context.

When to recall

Before starting related work or when context might exist:

• agent-carnet find — quick scan of summaries
• agent-carnet list — browse a folder. Each entry's tail shows (last_used: …, use_count: N) — a high use_count means the note has been load-bearing across past sessions
• agent-carnet read — actually read (bumps last_used; only use when the content matters)

Use use_count as one importance signal among several: combine with the summary, tags, and last_used before deciding what to read. A high-count note is not always relevant to the current question, and a freshly-saved load-bearing note starts at use_count: 0. agent-carnet list --sort use_count surfaces the most-applied notes when that ranking is the right lens.

When to call used

Call agent-carnet used after a carnet actually shaped your work:

• You applied the recorded fix and it solved the bug.
• You consulted the carnet before retrying a hypothesis and skipped a dead-end.
• You used the canonical name from a vocab carnet in new code instead of inventing your own.

Each used call bumps last_used AND increments use_count by one — a durable importance signal that survives across sessions and lets future readers (and agent-carnet list --sort use_count) surface load-bearing notes. Repeated calls each increment, so call once per real application event, not per session.

Reading a carnet does NOT count. read already keeps it alive (weak signal: bumps last_used only); used records that the note was worth keeping for a real reason (strong signal: bumps both, and +1s the use_count that future agents see in list output).

Hard rules

• --summary is required. Make it decisive — reading the summary in isolation tells the next reader (or the next agent) whether to read further.
• --agent claude-code is required.
• find does NOT bump anything. read bumps last_used. used bumps last_used AND increments use_count.
• updated tracks content modification only (save, save --update). It is independent of last_used and is not the lifespan driver.
• The 30-day expiry is automatic — do not manually clean up. keep: true pins permanent notes.
• Auto-prune runs on every CLI invocation; deleted carnets land in .carnet/.trash/ for 7 days before hard delete.

Path conventions

• / — kebab-case, no leading slash, no ...
• Categories are folders; create new ones freely as needed.
• Subcategories are allowed: deps/esm/iconv-issue works.

When to read references/

This SKILL.md is enough for everyday note-keeping. Open the references/ files only when one of these specific cases applies — they are not always-on context, so do not load them speculatively.

| Read this file | When | |---|---| | references/cookbook.md | You are about to use (or are being asked about) a tag-based pattern such as tags: [vocab] for project terminology or tags: [hypothesis] for debugging dead-ends. The file shows the full pattern, including how to structure the body and meta: for that pattern. | | references/frontmatter.md | You need to write or read the meta: extension namespace, set a non-trivial lifespan / keep, or understand why an unfamiliar frontmatter field is or is not preserved on save. |

If neither case applies, do not read references/. The base of this file already covers daily save/find/read/used/move/rm flows.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yamadashy
• Source: yamadashy/agent-carnet
• License: MIT
• Homepage: https://www.npmjs.com/package/agent-carnet

Install and usage instructions live in the source repository linked above.