Memory promote

memory-promote

Primary mode: cron (Mon/Wed/Fri 22:00 via launchd, wrapped by skill-wrapper). Manual invocation accepted for one-off catch-ups or backfill.

This skill is the bridge between the hippocampus (daily files, ephemeral, 7-day window) and the cortex (wiki, durable, indexed by both QMD and gbrain). Items that recur, get explicit-promote signals, or score high in classifier land in wiki. Lower-confidence items queue or stay daily. Stale items drop on the next compact.

Out of scope: writing new bullets to daily (memory-capture), reducing aged daily files to thin records (memory-compact), enforcing bloat invariants (memory-bloat-audit).

0. Start run telemetry

PARENT_ARG=""
if [ -n "${OPENCLAW_RUN_ID:-}" ]; then
  PARENT_ARG="--parent-run-id $OPENCLAW_RUN_ID"
fi
RUN_ID=$(${OPENCLAW_HOME}/bin/skill-run start \
  --skill memory-promote \
  --agent <current-agent> \
  --trigger <cron|interactive> \
  $PARENT_ARG)
export OPENCLAW_RUN_ID="$RUN_ID"

1. Discover candidate items

Scan all daily memory files inside the 7-day active window for each agent the sweep covers (default: all agents listed in ${OPENCLAW_HOME}/config/agents.map, e.g. main, dev):

WINDOW_START=$(date -v-7d +%F)
for ws in workspace workspace-dev; do
  find "${OPENCLAW_HOME}/$ws/memory" -name "*.md" -newer /tmp/__cutoff 2>/dev/null
done

(Use touch -t $(date -v-7d +%Y%m%d0000) /tmp/__cutoff first.)

From each file:

• Parse <!-- mc:item --> lines (those are candidate units).
• Skip lines that are already pointers (start with → see [wiki/) — those are graduated.
• Group by section (Decisions, Key Lessons, Carry Forward, What Happened, Quick Wins).

2. Score each candidate

Confidence score (0.0–1.0) is the sum of these signals, clipped to [0, 1]:

Signal	Weight	Detection
Same-or-similar item appears in N ≥ 2 daily files	+0.30 per dup (cap +0.60)	Jaccard ≥ 0.7 on bullet body, ignoring dates
Section is Decisions or Key Lessons	+0.25	Header lookup
Section is Carry Forward and recurred ≥ 3 days	+0.25	Combine signals
Body contains a promotion verb: "decided", "principle", "rule", "always", "never", "lesson", "post-mortem", "rca"	+0.15	Case-insensitive substring match
Body length ≥ 200 chars	+0.10	Length proxy for substance
User explicit signal in trailing text: "[promote]", "(durable)", "ship to wiki"	+0.30	Marker tokens
Body is a short todo ("call X", "ping Y", "check Z")	-0.30	Strong negative — todos don't belong in wiki
Section is What Happened AND single-day	-0.20	Activity log noise

Final bucket:

• ≥ 0.85 → auto-promote
• 0.60–0.85 → queue
• < 0.60 → drop (leave in daily; will compact or expire)

3. Auto-promote (≥ 0.85)

For each high-confidence cluster (a candidate plus any duplicates across days):

1. Call memory-canonicalize with: cluster body, proposed category (per memory-capture/references/routing-table.md), proposed slug. Receive verdict (create-new | update | update-suggested) and shaped markdown.
2. Call wiki-write with the canonical shape — lands in workspace/wiki/<category>/<slug>.md and updates wiki/INDEX.md. For update verdict, append to existing entry and refresh _Last updated.
3. Register a gbrain page so the durable entry is visible to gbrain queries:

   GBRAIN_HOME=$HOME gbrain sync --source workspace-main-wiki
   

   (Source ID matches the wiki being written. Main wiki = workspace-main-wiki; specialist wikis use agents-<agent>-wiki.)
4. Replace the original daily-file bullet(s) with a pointer:

   - **[Title]** — promoted to wiki on YYYY-MM-DD → see [wiki/<category>/<slug>.md] <!-- mc:item -->
   

   Edit each daily file in place. Preserve the <!-- mc:item --> stamp.
5. Trigger QMD re-index so the new wiki entry is searchable immediately:

   ${OPENCLAW_HOME}/bin/qmd-sync --collections wiki-main,agents-<agent>-wiki
   

   (If qmd-sync doesn't exist yet, fall back to: XDG_CONFIG_HOME=... qmd update && qmd embed.)

4. Queue (0.60–0.85)

For each mid-confidence item, append to the per-agent promote queue:

${OPENCLAW_HOME}/<workspace>/memory/promote-queue.md

Entry format:

## YYYY-MM-DD — <one-line summary> (score: 0.72)
- Source: <daily-file path>:<section>
- Proposed category: <category>
- Proposed slug: <kebab-case>
- Why queued: <signal explanation>
- Body:
  > <verbatim bullet body>

The user reviews the queue (or a future memory-graph-link skill surfaces it) and either flips the item to durable (forces ≥ 0.85 on next sweep with a [promote] marker) or leaves it to drop.

5. Drop (< 0.60)

No write. Item stays in daily file until memory-compact rolls the file over after 7 days.

6. End run telemetry

${OPENCLAW_HOME}/bin/skill-run end "$RUN_ID" \
  --outcome <success|partial|failure> \
  --task-completion <true|false> \
  --error-recovery true \
  --exit-code 0 \
  --completion-evidence "promoted: N, queued: M, dropped: K; wiki entries written: ...; gbrain sync: ok; qmd reindex: ok"

Outcome rubric:

• success — sweep completed for all covered workspaces, all auto-promotes landed in wiki AND gbrain AND QMD, queue updated, pointer-replacements in daily files succeeded.
• partial — one or more workspaces failed but others completed; or wiki write succeeded but gbrain/QMD sync failed (recoverable on next sweep).
• failure — total sweep aborted before any writes, OR wiki write succeeded but daily-file pointer replacement failed (leaves the duplicate the architecture forbids).

Why this skill exists

Without promotion, daily files become a graveyard:

• Same item carried forward 14 days = 14 copies of the same text → bloat + rank inversion in gbrain.
• Long-form decisions stuck in daily logs → never indexed as durable knowledge → re-derived from scratch later.
• No flow from capture → curation → durable means the system can't compound.

memory-promote is the consolidation phase. Together with memory-capture (write), memory-compact (prune), memory-bloat-audit (enforce), and memory-retrieve (read), it forms a closed loop where memory actually compounds.

Reference

• Architecture: cold-stored at workspace/cold-storage/wiki-projects/memory-system-redesign.md
• Shape helper: ${OPENCLAW_HOME}/skills/memory-canonicalize/SKILL.md
• Wiki writer: workspace/skills/wiki-write/SKILL.md
• gbrain page registration: ${OPENCLAW_HOME}/skills/brain-ops/SKILL.md
• QMD reindex: ${OPENCLAW_HOME}/skills/qmd/SKILL.md
• Per-agent routing: ${OPENCLAW_HOME}/skills/memory-capture/references/routing-table.md