/memory-team — Share Learnings via a Git-Tracked Artifact

Wires up memory.team_artifact: a diffable directory (.moflo/shared/learnings/, one JSON file per entry) that lives in the repo, reviewed like any other change. One file per entry means two PRs that each add a learning never conflict — on GitHub or locally. Teammates git pull it and their next session auto-imports the merged learnings; each entry is stamped with its git author. This skill also installs a pre-commit hook so the artifact is re-exported and staged on every commit — the learnings ride along in each PR with zero effort.

One mechanism, two audiences. A team and one person across machines are the same problem — push the durable slice through git. This skill covers both; the only difference is whether anyone else is on the other end. (Prefer a private, non-git file for your own machines? See "Advanced" at the bottom — flo memory sync moves a binary artifact that keeps embeddings verbatim.)

Arguments: $ARGUMENTS — pass --solo to skip team-oriented wording (still identical setup).

When to use

• "share learnings / memory with my team"
• "commit our moflo knowledge to the repo"
• "sync memory between my laptop and desktop"
• "auto-export the learnings on every PR / commit"
• "set up team_artifact"
• "learnings.jsonl conflicts on every PR" → Step 2b (it migrates itself)

For many worktrees on one machine, that's a shared local store, not git — use /memory-worktree.

Procedure

Step 0 — Memory first (gate requirement)

mcp__moflo__memory_search { query: "team_artifact learnings jsonl git share import session-start", namespace: "guidance" }

Step 1 — Confirm this is the git path

• Team on one repo, or one person across machines who both push to the same repo? → correct skill.
• Many worktrees on one machine, no cross-machine need? → redirect to /memory-worktree.
• A private synced file (Dropbox/USB), no git? → point at the Advanced section (flo memory sync).

Step 2 — Set the config

Read the project-root moflo.yaml and add/merge under memory: — preserve existing keys:

memory:
  team_artifact: .moflo/shared/learnings/

An existing team_artifact: .moflo/shared/learnings.jsonl line is fine as it is — a .jsonl path names the directory beside it, so there is nothing to change.

Step 2b — Existing learnings.jsonl setups migrate themselves

Do nothing. The old single file conflicted whenever two branches added neighbouring keys, and GitHub ignores merge drivers, so merge=union never helped on the PR page. The next team-export — normally the pre-commit hook — moves it:

Existing hooks keep working: when the artifact is git-tracked, team-export stages it itself. Refresh an old hook block to the one in Step 4 anyway, so it names the directory.

Step 3 — First export (creates + tracks the artifact)

npx flo memory team-export 2>&1

This writes the artifact, pins it to LF in .gitattributes, and rewrites .gitignore from a blanket .moflo/ to .moflo/* + !/.moflo/shared/ so the shared artifact is tracked while the machine-local DB stays ignored. Surface what it reports (entries added, .gitignore change). If there are no durable learnings yet, that's fine — the artifact is created empty and fills in as learnings accrue.

Step 4 — Install the pre-commit hook

The hook re-exports and stages the artifact on every commit, so it's never stale in a PR. Handle three cases; do not clobber an existing hook.

Define the managed block (POSIX sh — git runs hooks via sh on Linux, macOS, and Windows):

# >>> moflo team-export (managed by /memory-team) >>>
# Refresh the shared learnings artifact so it rides along in this commit.
npx --no-install flo memory team-export >/dev/null 2>&1 || true
if [ -n "$(git status --porcelain -- .moflo/shared/learnings 2>/dev/null)" ]; then
  git add -A -- .moflo/shared/learnings
fi
# <<< moflo team-export (managed by /memory-team) <<<

Install it:

1. Husky present (.husky/pre-commit exists) → append the block to that file.
2. Existing .git/hooks/pre-commit → if it already contains the moflo team-export markers, replace the block between them with the one above (an older block names learnings.jsonl); otherwise append the block after the shebang.
3. No hook → create .git/hooks/pre-commit with #!/bin/sh as the first line, then the block.

On POSIX, make it executable:

node -e "require('fs').chmodSync('.git/hooks/pre-commit', 0o755)"

(Use the Node fs.chmodSync primitive, not chmod — it's a harmless no-op on Windows. Never shell out to chmod in a cross-platform step.)

• The || true keeps a failed export from ever blocking a commit.
• The git status --porcelain guard means the artifact is only re-staged when it actually changed — no empty-diff churn on unrelated commits. -A also stages entry files export removed (expired tombstones).
• Using a custom team_artifact path? Substitute it for .moflo/shared/learnings in both lines.

Step 5 — Commit and share

git add .moflo/shared/learnings/ .gitignore .gitattributes
git commit -m "chore: share moflo learnings"
git push

Teammates (or your other machine) run git pull; their next session auto-imports the merged artifact — no manual import command. To pull learnings without waiting for a restart:

npx flo memory team-import 2>&1

Output

End with a single-block summary:

Team Memory Sharing — configured
─────────────────────────────────
Artifact:    .moflo/shared/learnings/  (git-tracked, one file per entry, author-stamped)
Config key:  memory.team_artifact
PR hook:     .git/hooks/pre-commit  → re-exports + stages on every commit
Import:      automatic at session start (after git pull)  |  npx flo memory team-import
Scope:       learnings + knowledge only (embeddings regenerated on import)

Advanced — one person, a private file (no git)

If you're solo and would rather not commit learnings, carry a binary artifact instead. It keeps the embeddings verbatim (no recompute on the far side) and merges idempotently:

npx flo memory sync --to   ~/Dropbox/moflo-learnings.db   # machine A
npx flo memory sync --from ~/Dropbox/moflo-learnings.db   # machine B

Same durable slice, no repo footprint, no review — the trade-off is it's not diffable and not shared.

Rules

• Don't reimplement export/import/migration or the .gitignore rewrite — flo memory team-export and team-import own all of it. This skill sets config, runs them, and installs the hook.
• Don't add a merge=union attribute for the artifact — GitHub ignores it, and the directory layout does not need it.
• Don't overwrite an existing pre-commit hook or husky config — append the marked block only.
• Don't use chmod / mkdir / rm in shell steps — Windows lacks them on PATH; use Node fs.
• Embeddings are intentionally omitted from the artifact (384 floats/entry would wreck diffs) — they're regenerated on import. Don't try to add them.

See Also

• /memory-worktree — same-machine sharing across worktrees / Conductor (a shared local store, not git)
• flo memory team-export --help / flo memory sync --help
• docs/blog/memory-sharing.html — the design rationale (durable slice, why not share the whole DB)