memarium

/memarium — digest sessions into the typed Memory OS

This skill walks the in-session Claude (you) through digesting the user's already-synced AI coding sessions into ONE per-project knowledge layer: typed memory — episodic (one per work thread: the arc, dead-ends, decisions), semantic (durable facts), procedural (how-to playbooks + gotchas), core (never-forget rules) — plus the two derived layers, the entity wiki and distilled Q&A. It is AI-native: agents read this via /memarium-recall and the SessionStart primer. There is no human "book" — no chronicles, no topics, no reading site.

Pure mechanical CLI handles I/O ("$VBP" prepare / "$VBP" memory-write / "$VBP" memory-propose / "$VBP" entity-write / "$VBP" qa-write / "$VBP" skip-write / "$VBP" finalize / "$VBP" list-projects); the LLM work — segmentation + writing — is yours, in this conversation, with full context.

Inputs you assume

• The plugin's orchestrate step (Pre-step + Step 0 below) scans the user's local Claude Code + Copilot Chat + Codex Desktop/interactive CLI JSONL, writes rendered .md into ~/.memarium/session-repo/raw_sessions/..., and updates .memarium/index.json. You don't need to check whether sync ran — the plugin scans on every run.
• If the user also installed the optional memarium npm CLI for cross-device sync, fine — both write the same spool. Don't gate on it; it's not required.

---

Step −1 — Locate the plugin binary (DO THIS FIRST OF ALL)

${CLAUDE_PLUGIN_ROOT} is not populated in your in-session Bash environment, so discover the plugin's bin path and stash it in a shell variable:

VBP="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/bin/memarium-plugin.js}"
[ -x "$VBP" ] || VBP=$(ls -d ~/.claude/plugins/cache/*/memarium/*/bin/memarium-plugin.js 2>/dev/null | awk -F/ '{print $(NF-2)"\t"$0}' | sort -V | tail -1 | cut -f2-)
echo "VBP=$VBP"

Confirm it resolves (picks the highest semver version):

[ -x "$VBP" ] && "$VBP" --version