memory-upgrade
Diagnose and fix broken memory search in OpenClaw. Enables local embeddings, hybrid search (BM25+vector), session transcript indexing, MMR diversity, and temporal decay — all running locally with zero API keys. Use when: memory_search returns empty results, agent has poor cross-session recall, user wants to upgrade their memory system, or after a fresh OpenClaw install.
- License 许可
- MIT
- 最近更新
- 2026-09-02
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- memory-upgrade
- description
- Diagnose and fix broken memory search in OpenClaw. Enables local embeddings, hybrid search (BM25+vector), session transcript indexing, MMR diversity, and temporal decay — all running locally with zero API keys. Use when: memory_search returns empty results, agent has poor cross-session recall, user wants to upgrade their memory system, or after a fresh OpenClaw install.
02 资源文件 5
- 脚本scripts/configure.sh3.4 KB
- 脚本scripts/diagnose.sh3.1 KB
- 脚本scripts/organize.sh2.5 KB
- 脚本scripts/tag.sh3.7 KB
- 脚本scripts/verify.sh4.4 KB
03 SKILL.md 正文
Memory Upgrade
Most OpenClaw installs have broken memory search — the memory_search tool returns empty results because no embedding provider is configured. OpenClaw auto-detects OpenAI → Google → Voyage keys; if none exist, embeddings stay disabled silently.
This skill fixes it with fully local inference. No API keys. No data leaves the machine.
Quick Start
# 1. Diagnose
bash scripts/diagnose.sh
# 2. Fix (patches openclaw.json, restart [REDACTED] after)
bash scripts/configure.sh
# 3. Restart [REDACTED]
openclaw [REDACTED] restart
# 4. Verify (waits for indexing, runs test query)
bash scripts/verify.sh
Optional Enhancements
# Organize memory files into clean directory structure
bash scripts/organize.sh
# Add YAML frontmatter tags to untagged files
bash scripts/tag.sh
What Gets Enabled
| Feature | Details |
|---|---|
| Local embeddings | embeddinggemma-300m (~328MB GGUF, auto-downloads) |
| Hybrid search | BM25 keyword + vector semantic (70/30 weight) |
| Session transcripts | Past conversations become searchable |
| MMR diversity | Reduces duplicate/overlapping results (λ=0.7) |
| Temporal decay | Recent memories rank higher (30-day half-life) |
| Embedding cache | 50k entries, avoids re-embedding unchanged text |
| File watcher | Auto-reindexes when memory files change |
How It Works
- Patches
agents.defaults.memorySearchinopenclaw.json - Uses
node-llama-cpp(ships with OpenClaw) for local embeddings - Vector search via
sqlite-vec(ships with OpenClaw) - No external dependencies required
Notes
- First search after restart may be slow (model loads into memory)
- Initial indexing takes 30-120s depending on file count
- Embedding model runs on CPU (ARM/x86), ~768-dim vectors
- Compatible with existing memory files — no migration needed
04 同仓库更多 Skills
three-shifts
Cyclic shift execution engine. Plans tasks 3x daily (6 AM, 2 PM, 10 PM), decomposes them into granular steps, then executes via 15-minute cron cycles. Each cycle reads state files, picks the next step, executes it, writes results back. Errors are logged and skipped — never fatal. Planning uses Claude 4.6; execution uses GLM-5.
xmtp-comms-guard
Security middleware for all XMTP communications in EverClaw. Enforces guarded client usage with validation, integrity checks, and fail-closed security policies. Integrates approval flows for sensitive operations. Use when integrating XMTP messaging, configuring communication security, or auditing guarded client enforcement.
three-shifts
Daily standup engine. Plans tasks 3x daily (6 AM, 2 PM, 10 PM) and delivers them for approval. Execution happens in the main session via direct conversation. Night shifts auto-approve carryover from earlier in the day.
super-helper
A helpful utility skill for agents
github-issues
Fetch and manage GitHub issues via the API
code-formatter
Format and lint code files