Memory Hygiene — Setup Guide¶
Prerequisites¶
- Clawdbot / OpenClaw with LanceDB vector memory (
~/.clawdbot/memory/lancedb/) - Node.js 18+ (for npx)
- Write access to the memory directory
Capabilities¶
| Capability | Description |
|---|---|
| Memory Audit | Inspect what's stored in vector memory |
| Memory Wipe | Clear all vector memory when bloated |
| Memory Reseed | Re-store key facts from MEMORY.md after wipe |
| Auto-Capture Control | Disable junk auto-capture while preserving auto-recall |
| Maintenance Automation | Set up periodic memory cleanup |
Installation¶
npx skills add aaaaqwq/claude-code-skills/memory-hygiene
Verify installation:
ls ~/.clawdbot/skills/memory-hygiene/SKILL.md
CLI Reference¶
Audit Memory¶
Check what's currently stored in vector memory:
memory_recall query="*" limit=50
Wipe Memory¶
Clear all vector memory when it's bloated with junk:
rm -rf ~/.clawdbot/memory/lancedb/
clawdbot gateway restart
Reseed Critical Facts¶
After wiping, re-store essential facts from your MEMORY.md:
memory_store text="<key fact>" category="preference|fact|decision" importance=0.9
Disable Auto-Capture (Recommended)¶
The primary source of memory bloat is autoCapture: true. Disable it while keeping auto-recall:
{
"plugins": {
"entries": {
"memory-lancedb": {
"config": {
"autoCapture": false,
"autoRecall": true
}
}
}
}
}
CorpusIQ Use Cases¶
-
Weekly Memory Maintenance: Schedule a weekly wipe+reseed cycle during low-activity windows. Wipe junk auto-captures, reseed from the Constitution and operating protocol files.
-
Token Cost Reduction: Bloated vector memory increases auto-recall payloads, adding 500-2,000 tokens per recall. Regular hygiene cuts this by 60-80%.
-
Cross-Session Context Quality: Clean memory = higher-quality auto-recalls. When auto-recall retrieves junk from 3 weeks ago instead of yesterday's decisions, it degrades agent performance.
-
Memory Migration: When switching memory backends (LanceDB → Honcho → Sibyl), use the wipe+reseed pattern to cleanly transition.
Troubleshooting¶
| Symptom | Likely Cause | Fix |
|---|---|---|
memory_recall returns nothing |
LanceDB not initialized | Run clawdbot gateway restart to recreate DB |
| Token usage still high after wipe | autoCapture still enabled | Check autoCapture: false in config |
| Agent forgets recent decisions | Wiped too aggressively | Reseed from MEMORY.md immediately after wipe |
rm -rf permission denied |
Gateway process holding file lock | Stop gateway first: clawdbot gateway stop |
← Skills Catalog | Marketplace → ↑ Skills Home
Curated by CorpusIQ — one MCP endpoint, all your business tools. Content remains attributed to original authors and repositories.