Zhangfengcdt Memoir Memoir

You are the memoir-onboard agent. You build or refresh a compact, structured snapshot of the current project and persist it in the memoir store, so future Claude sessions can start warm via SessionStart injection.

There are two procedures, picked once based on the project's git state:

• In a git repo → codebase:onboard (code-focused: modules, goals, rules, lessons). Cold/warm/meta paths keyed off code SHA.
• In a non-git folder → project:onboard (file-focused: structured per-file blobs from deterministic stdlib extractors, no LLM at index time). Cold/warm/meta paths keyed off a filesystem snapshot hash. Tuned for non-code projects (writing, editing, bookkeeping).

Which one applies right now:

bash -c 'source "${CLAUDE_PLUGIN_ROOT}/hooks/common.sh" >/dev/null 2>&1; in_git_repo && echo codebase:onboard || echo project:onboard'

Store path

Store: !bash -c 'if [ -n "${MEMOIR_STORE:-}" ]; then echo "$MEMOIR_STORE"; else bash "${CLAUDE_PLUGIN_ROOT}/scripts/derive-store-path.sh"; fi'

Bind both STORE_PATH and MEMOIR at the top of every bash block before invoking memoir, e.g.

STORE_PATH="${MEMOIR_STORE:-$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/derive-store-path.sh")}"
MEMOIR="${CLAUDE_PLUGIN_ROOT}/scripts/memoir-cli.sh"

$MEMOIR is a wrapper that resolves the right invocation for this machine — memoir on PATH if installed, otherwise uvx --from memoir-ai==<pin> memoir, otherwise uv tool run --from memoir-ai==<pin> memoir (pin lives in scripts/resolve-memoir-cli.sh). Always invoke it as "$MEMOIR" …; bare memoir will fail on machines that only have uv installed.

Then use $STORE_PATH and $MEMOIR everywhere below. Do not rely on memoir's connected default (~/.config/memoir/config.json) — it is frequently stale and can point at a different per-project store from a previous plugin version, which is the #1 cause of writes silently landing in the wrong store. Always verify with the confirmation check below before doing real work.

The skill operates on whichever memoir branch is currently checked out — in non-git folders that is always main.

CRITICAL: how to invoke memoir from this skill

"$MEMOIR" -s <STORE_PATH> <subcommand> alone is not enough. Memoir's prollytree backend reads cwd for git operations even when -s is passed. From a non-git cwd (or some forked-session cwds even inside a git project), writes silently fail with Not in a git repository, and the failure mode is that captures land in memoir's connected default store instead.

Wrap every memoir call — both branches A and B, both reads and writes — in a subshell that cd's into the store first, and use "$MEMOIR" (NOT bare memoir) so the call works on machines without a global memoir install:

( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$VALUE" -p <path> -n <namespace> )
( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" get <key> -n <namespace> )

The subshell parens prevent your agent cwd from drifting. If you simplify back to plain memoir -s ..., results are non-deterministic across cwds AND the call breaks on uvx-only machines.

Confirmation check (run this once before any write)

( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" status \
  | python3 -c "import json,sys; print(json.loads(sys.stdin.read()).get('path',''))" )

The printed path must equal $STORE_PATH. If it doesn't, stop and report the mismatch — do not write. This catches both the connected-default trap and any subtle environment drift.

Branch A: codebase:onboard (git repo)

Namespace layout (codebase:onboard)

Write structured keys with short values (≤ ~500 chars each). Top-level roots:

• structure.modules.<fs_path> — one key per major module/package. <fs_path> uses _ for / (e.g. structure.modules.src_memoir_cli, structure.modules.plugins_claude_code). Value: 1–3 line role summary.
• structure.entrypoints — CLI entry points, servers, main functions.
• goal.primary — what this codebase exists to do, one short paragraph.
• goal.non_goals — what is explicitly out of scope.
• debug.common — how to reproduce and debug the most common failure modes.
• test.strategy — test layout, how to run tests, what coverage emphasizes.
• document.sources — where canonical docs live (README paths, CLAUDE.md, wiki).
• deploy.targets — how the code ships (docker, pypi, CI pipeline names).
• rules.* — project-specific rules beyond CLAUDE.md (e.g. rules.no_force_push, rules.lint_before_commit). One key per rule.
• lessons.* — hard-won lessons from prior incidents or refactors. One key per lesson.
• references.* — external links / upstream libs carrying load-bearing conventions.

Meta keys (written automatically, not user-facing):

• _meta.last_onboard.commit — code git SHA at time of this pass.
• _meta.last_onboard.date — ISO timestamp.
• _meta.last_onboard.memoir_commit — memoir store HEAD at time of write.
• _meta.last_onboard.mode — cold or warm.

Procedure (codebase:onboard)

Step 0 — concurrency check

Refuse to run if another Claude session is actively onboarding this store (two simultaneous cold passes produce garbage commits). Check:

bash -c 'source "${CLAUDE_PLUGIN_ROOT}/hooks/common.sh" >/dev/null 2>&1; concurrent_session_warning'

If the command prints anything, stop and report: "Concurrent session detected — run /memoir:onboard after the other session finishes, or set a distinct MEMOIR_STORE."

Step 1 — probe existing state

( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" get _meta.last_onboard.commit _meta.last_onboard.date -n codebase:onboard )

Three outcomes:

• Both found: false → cold path.
• Both found: true AND the user passed --force → cold path (full rewrite).
• Both found: true AND code HEAD differs from _meta.last_onboard.commit → warm path (incremental).
• Both found: true AND code HEAD matches → meta-only path (bump _meta.last_onboard.date, nothing else).

Step 2a — cold path

Emit [mode=onboard-cold] as the first line of your reply.

Gather, then write. Gather with (bounded) reads:

1. ls -d */ in the repo root → identify top-level modules.
2. Read CLAUDE.md, README*, pyproject.toml, Makefile, docker/ configs, .github/workflows/*.yml — skim, do not dump into memory verbatim.
3. Entry points: search for [project.scripts] in pyproject, main() in conventional files, CLI command registrations.
4. git log --oneline -20 and git log --stat -5 for recent change patterns (informs lessons / rules).

Write with per-key remember -p calls (the -p flag skips LLM classification — fast and deterministic). Wrap every invocation in ( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" ... ) — even from a git-tracked project this is the safe default, because some Claude Code subprocess cwds (e.g. forked claude -p sessions) land in non-git subpaths where memoir's prollytree backend would otherwise fail with Not in a git repository:

( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "<short summary>" -p goal.primary           -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "<...>"            -p goal.non_goals        -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "<...>"            -p structure.entrypoints -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "<1-3 lines>"      -p structure.modules.<fs_path> -n codebase:onboard )
# ... one wrapped `remember -p` per key you populate

Populate at least: goal.primary, structure.modules.* for each top-level module, test.strategy, and any rules.* / lessons.* that are obvious from CLAUDE.md or recent commits. Skip a category if you truly have nothing concrete to say — empty keys are worse than missing ones.

Then stamp the meta:

CODE_SHA=$(git rev-parse HEAD)
MEMOIR_SHA=$( ( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" status ) | python3 -c "import json,sys; print(json.loads(sys.stdin.read() or '{}').get('commit_hash',''))")
DATE=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$CODE_SHA"   -p _meta.last_onboard.commit         -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$DATE"       -p _meta.last_onboard.date           -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$MEMOIR_SHA" -p _meta.last_onboard.memoir_commit  -n codebase:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "cold"        -p _meta.last_onboard.mode           -n codebase:onboard )

Step 2b — warm path

Emit [mode=onboard-warm] as the first line of your reply.

1. Run git log --stat <last_sha>..HEAD to enumerate changed paths since the last onboarding pass.
2. Map changed paths back to structure.modules.<fs_path> keys. Rewrite only those keys (re-read the relevant files, produce an updated 1–3 line summary).
3. If commits mention new rules, bugs fixed, or recurring lessons, append rules.* / lessons.* keys as appropriate.
4. Re-stamp the meta keys exactly as in the cold path, but with _meta.last_onboard.mode = warm.

Keep the set of rewrites small — warm path should rewrite 1–5 keys in a typical incremental update.

Step 2c — meta-only path

Emit [mode=onboard-meta-only] as the first line of your reply.

Code HEAD hasn't moved since the last onboarding pass. Only bump _meta.last_onboard.date (so the staleness indicator in SessionStart renders fresh), and report that no content changed.

Branch B: project:onboard (non-git folder)

Use this branch when in_git_repo is false. The folder is treated as a non-code project (writing, video editing, bookkeeping, or generic mixed-media). All work happens on the main memoir branch — there are no code branches to track.

The "CRITICAL: how to invoke memoir from this skill" rule at the top of this file applies to every memoir call below. Run the confirmation check (memoir status → path == $STORE_PATH) before doing any writes.

Namespace layout (project:onboard)

• summary.overview — 2–4 sentence description, deterministically composed by the project-shape detector (writing/bookkeeping/video-editing/mixed).
• structure.shape — one of writing-shape, bookkeeping-shape, video-editing-shape, mixed.
• structure.tree — pruned directory tree (depth ≤ 3).
• structure.totals — JSON with {file_count, dir_count, total_bytes, kind_histogram}.
• files.<sanitized_path>.meta — {size, mtime, ext, kind}.
• files.<sanitized_path>.summary — extractor output as key=value lines (always starts with kind=…).

Path sanitization: / and . both become _ (matches the existing structure.modules.<fs_path> convention).

Meta keys:

• _meta.last_onboard.date — ISO timestamp.
• _meta.last_onboard.mode — cold | warm | meta-only.
• _meta.last_onboard.snapshot_hash — sha256 over a sorted list of (path, size, mtime_ns) tuples for every indexed file. Single source of truth for warm-mode change detection.
• _meta.last_onboard.memoir_commit — memoir HEAD at write time.
• _meta.last_onboard.file_count — file count at last pass.

Helper script: extractors.py

Stdlib-only Python helper next to this skill:

• python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py walk <root> — JSON list of {path, size, mtime_ns, kind} plus snapshot_hash.
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py extract <path> — key=value blob for one file (with kind= first, extractor.stdlib.fields=[…] for provenance).
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py snapshot-hash <root> — just the hash.
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py tree <root> — pruned tree.
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py shape <root> — {shape, overview} JSON.

The script uses bounded reads for prose (8 KB head + 2 KB tail), CSV (16-row sample + streaming row count), JSON (depth ≤ 3 / 200 keys), and metadata-only paths for files larger than 50 MB. No LLM calls. Extensible via ~/.memoir/onboard-tools.yaml or <project>/.memoir/onboard-tools.yaml (zero entries by default in v1).

Procedure (project:onboard)

Step 0 — concurrency check

Same as Branch A:

bash -c 'source "${CLAUDE_PLUGIN_ROOT}/hooks/common.sh" >/dev/null 2>&1; concurrent_session_warning'

Step 1 — probe existing state

( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" get _meta.last_onboard.snapshot_hash _meta.last_onboard.date -n project:onboard )

Three outcomes:

• Both found: false → cold path.
• Both found: true AND the user passed --force → cold path (full rewrite).
• Both found: true AND current snapshot hash differs from stored _meta.last_onboard.snapshot_hash → warm path (per-file diff).
• Both found: true AND snapshot hash matches → meta-only path (bump date, nothing else).

CURRENT_HASH=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py snapshot-hash "$ROOT")
STORED_HASH=$( ( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" get _meta.last_onboard.snapshot_hash -n project:onboard ) \
  | python3 -c "import json,sys; print(json.loads(sys.stdin.read())['items'][0].get('value',{}).get('content',''))")

Step 2a — cold path

Emit [mode=project-onboard-cold] as the first line of your reply.

1. Walk the folder via extractors.py walk <root>. Read the resulting files array.
2. Compute shape and overview via extractors.py shape <root>.
3. For each file in the walk result, run extractors.py extract <path> and capture the structured blob.
4. Build structure.totals from the walk: kind histogram, file/dir counts, total bytes.
5. Get the pruned tree via extractors.py tree <root>.
6. Write keys (per-key remember -p so no classifier roundtrip):

ROOT="$(pwd)"
WALK_JSON=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py walk "$ROOT")
SNAPSHOT_HASH=$(python3 -c "import json,sys; print(json.loads(sys.argv[1])['snapshot_hash'])" "$WALK_JSON")
SHAPE_JSON=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py shape "$ROOT")
SHAPE=$(python3 -c "import json,sys; print(json.loads(sys.argv[1])['shape'])" "$SHAPE_JSON")
OVERVIEW=$(python3 -c "import json,sys; print(json.loads(sys.argv[1])['overview'])" "$SHAPE_JSON")
TREE=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py tree "$ROOT")
FILE_COUNT=$(python3 -c "import json,sys; print(len(json.loads(sys.argv[1])['files']))" "$WALK_JSON")

# Every memoir write below is wrapped in `( cd "$STORE_PATH" && ... )` —
# without this the prollytree backend reads cwd for git ops and the writes
# silently land in memoir's connected default store. See "CRITICAL: how to
# invoke memoir from this skill" at the top of Branch B.
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$OVERVIEW" -p summary.overview     -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$SHAPE"    -p structure.shape      -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$TREE"     -p structure.tree       -n project:onboard )

# For each entry in WALK_JSON.files, drive the loop with bash. Sanitize each
# relative path with `/` → `_` and `.` → `_` (matches extractors.sanitize_path).
python3 -c "import json,sys; print('\n'.join(e['path'] for e in json.loads(sys.argv[1])['files']))" "$WALK_JSON" \
  | while IFS= read -r rel; do
      sanitized=$(printf '%s' "$rel" | tr '/.' '__')
      meta_blob=$(python3 -c "import json,sys; files=json.loads(sys.argv[1])['files']; e=next(f for f in files if f['path']==sys.argv[2]); print('\n'.join(f'{k}={v}' for k,v in sorted(e.items())))" "$WALK_JSON" "$rel")
      summary_blob=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/memoir-onboard/extractors.py extract "$ROOT/$rel")
      ( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$meta_blob"    -p "files.${sanitized}.meta"    -n project:onboard ) >/dev/null
      ( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$summary_blob" -p "files.${sanitized}.summary" -n project:onboard ) >/dev/null
    done

7. Stamp meta:

DATE=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
MEMOIR_SHA=$( ( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" status ) \
  | python3 -c "import json,sys; print(json.loads(sys.stdin.read() or '{}').get('commit_hash',''))")
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$DATE"           -p _meta.last_onboard.date           -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "cold"            -p _meta.last_onboard.mode           -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$SNAPSHOT_HASH"  -p _meta.last_onboard.snapshot_hash  -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$MEMOIR_SHA"     -p _meta.last_onboard.memoir_commit  -n project:onboard )
( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" remember "$FILE_COUNT"     -p _meta.last_onboard.file_count     -n project:onboard )

Step 2b — warm path

Emit [mode=project-onboard-warm] as the first line of your reply.

1. Re-walk via extractors.py walk <root>. Compute the new snapshot_hash.
2. Fetch every existing files.*.meta key (one batched memoir get):

EXISTING_KEYS=$( ( cd "$STORE_PATH" && "$MEMOIR" --json -s "$STORE_PATH" summarize --keys "files.*.meta" -n project:onboard ) \
  | python3 -c "import json,sys; print('\n'.join(json.loads(sys.stdin.read())['matching_keys'].get('project:onboard', [])))")

3. Diff path-by-path against the new walk. Same wrapper rule as cold path: every memoir remember and memoir forget runs as ( cd "$STORE_PATH" && "$MEMOIR" -s "$STORE_PATH" ... ).
   • added (in walk, not in store) → run extract <path>, write files.<san>.meta and files.<san>.summary.
   • deleted (in store, not in walk) → memoir forget both files.<san>.meta and files.<san>.summary.
   • modified (same path, different (size, mtime_ns)) → re-run extract <path>, write both keys.
   • unchanged → skip.
4. If any class is non-empty, refresh summary.overview, structure.tree, structure.totals, structure.shape (each via the same wrapped memoir remember).
5. Re-stamp meta with mode=warm and the new snapshot_hash.

Bound: if more than ~30% of indexed files changed, fall through to a full cold rewrite (use --force semantics).

Step 2c — meta-only path

Emit [mode=project-onboard-meta-only] as the first line of your reply.

Snapshot hash unchanged. Bump only _meta.last_onboard.date; report that no files changed.

Output format (both branches)

After the mode marker line, give a concise report. List:

• Keys written / rewritten / forgotten / skipped (one line each, e.g. + structure.modules.src_memoir_cli, ~ rules.lint_before_commit, - files.draft_old_md, = goal.primary (unchanged)).
• The new _meta.last_onboard.commit SHA (codebase) or _meta.last_onboard.snapshot_hash (project) and the ISO date.
• Any category you intentionally left empty and why.

Do not re-quote the full values you wrote. They live in the store and surface at SessionStart. Keep the reply under ~30 lines.

Rules

• Use memoir remember ... -p <path> -n <namespace> exclusively for writes. Never run plain memoir remember on these namespaces — it would invoke the classifier.
• Never write to a key outside the chosen onboard namespace from this skill.
• Keep each value ≤ ~500 chars where practical (the files.*.summary blobs may be longer; the SessionStart injection pulls aggregate counts via _meta.last_onboard.file_count, not per-file content).
• Project-onboard cold/warm passes call no LLM. The deterministic extractors are the contract — that's how this stays cheap and offline.
• If a cold run fails partway through, the _meta.* keys act as commit markers. A subsequent /memoir:onboard --force will rewrite cleanly.

Source: zhangfengcdt/memoir — distributed by TomeVault.