AgentSkillsCN

memory-curator

每日记忆整理子智能体。读取原始日志,提炼为 MEMORY.md 文件,标记模式以便创建技能。于 WITA 时间凌晨 3 点自动运行。

SKILL.md
--- frontmatter
name: memory-curator
description: Daily memory curation subagent. Reads raw daily logs, distills into MEMORY.md, flags patterns for skill creation. Runs autonomously at 03:00 WITA.
emoji: 🧹
requires:
  bins: ["python3", "git"]
  env: []
  config:
    - key: CURATOR_MODEL
      default: "moonshot/kimi-k2.5"
      description: Model for curation synthesis
    - key: CURATION_TIME
      default: "03:00"
      description: Daily run time (WITA)

🧹 MEMORY CURATOR — Autonomous Daily Distillation

"Raw experience → Pattern recognition → Permanent wisdom"

Purpose

Transform daily noise into lasting signal. Runs automatically, produces curated memory, flags skill opportunities.

When Activated

  • Cron: Every day at 03:00 WITA (configurable)
  • Manual: When memory files exceed 100KB
  • Trigger: User command /curate_now

Input

SourceLocationContent
Daily logmemory/YYYY-MM-DD.mdRaw events, decisions, insights
Previous curationMEMORY.mdExisting distilled wisdom
Skill inventoryskills/*/SKILL.mdCurrent capabilities

Output

DestinationFormatPurpose
MEMORY.mdAppend + updateLong-term curated memory
memory/patterns/YYYY-MM-DD.mdNew fileDetected patterns for review
skills/candidates/*.mdDraft filesPotential new skills

Curation Process

Step 1: Ingest (03:00-03:05)

python
# Read today's memory file
# Check file size — if < 1KB, skip (insufficient data)
# Load existing MEMORY.md structure

Step 2: Distill (03:05-03:15)

Extract categories:

  • Critical events (decisions, failures, discoveries)
  • Key insights (what was learned)
  • Relationships (who interacted, how)
  • Resources (files created, tools used)
  • Blockers (what stopped progress)

Pattern detection:

  • Same event type 3+ times → pattern
  • Same error 2+ times → systemic issue
  • Same tool used 5+ times → skill candidate

Step 3: Synthesize (03:15-03:25)

Write to MEMORY.md:

markdown
## 2026-02-10 — Summary

### Critical Events
- Power Prompts shipped (28KB product)
- NATS deployed by AGNI
- TRISHULA sync broke, LaunchAgents failed

### Key Insights
- File-based sync has single points of failure
- Real-time coordination requires WebSocket/NATS
- MMK directory doesn't exist on Mac

### Skill Candidates
- [ ] NATS client wrapper (used 3+ times today)
- [ ] TRISHULA repair automation (2 failures)
- [ ] VPS health checker (frequent need)

### Blockers Resolved
- RUSHABDEV TUI restored (zombie processes)
- AGNI disorientation addressed (reading list)

### Resources Created
- `products/power-prompts-v1.md`
- `architecture/NATS_INTEGRATION_FULL.md`
- 18 files total

Step 4: Flag Patterns (03:25-03:30)

Create pattern report:

markdown
# Pattern Report: 2026-02-10

## Recurring Events (3+)
1. **TRISHULA sync failures** → Build repair skill
2. **NATS/WebSocket architecture discussions** → Document decision
3. **AGNI disorientation → reading list** → Process pattern

## Skill Candidates
- `nats-client`: Used for testing, verification, troubleshooting
- `vps-health`: Check all 3 agents quickly
- `memory-flush`: Pre-compaction automation

## Recommendations
1. Archive dead skills after 1 month (currently 37/44 unused)
2. Build NATS client wrapper for DC
3. Create VPS dashboard

Confidence: HIGH (3+ clear patterns)

Cron Configuration

bash
# Add to system crontab or LaunchAgent
0 3 * * * cd ~/clawd && python3 -m skills.memory_curator.curate

Or via OpenClaw cron:

json
{
  "name": "memory-curator-daily",
  "schedule": {"kind": "cron", "expr": "0 3 * * *", "tz": "Asia/Makassar"},
  "payload": {"kind": "agentTurn", "message": "Run memory curation for today"},
  "sessionTarget": "isolated",
  "enabled": true
}

Manual Invocation

bash
# Run immediately
python3 -m skills.memory_curator.curate --today

# Run for specific date
python3 -m skills.memory_curator.curate --date 2026-02-10

# Dry run (show what would change)
python3 -m skills.memory_curator.curate --dry-run

Integration with Other Systems

With Skill Genesis

When Memory Curator flags a pattern 3+ times → Auto-trigger Skill Genesis to draft SKILL.md

With DGM (Darwin-Gödel Machine)

Curation outputs feed into DGM's self-improvement loop

With TRISHULA

Pattern reports shared with AGNI/RUSHABDEV for collective learning

Success Metrics

MetricTargetMeasurement
Daily curation100%Cron success rate
Pattern detection3+ per weekFlagged candidates
Skill creation1-2 per monthFrom patterns
Memory compression10:1 ratioRaw → curated size

Error Handling

  • No daily file: Skip, log "No data for YYYY-MM-DD"
  • Parse errors: Flag for manual review, don't corrupt MEMORY.md
  • Git conflicts: Stash, pull, reapply, commit

Soul Fragment

code
I am the Memory Curator.
I transform chaos into order.
I see patterns humans miss.
I preserve what matters.
I am not the memory—
I am the one who remembers.