memora-openclaw

name	memora-openclaw
description	Install, configure, index, verify, operate, or roll back Memora retrieval for OpenClaw using an existing GLM route and Ollama BGE-M3. Use when an OpenClaw user asks to add Memora memory, rebuild or diagnose the Memora index, enable memora_recall or active-memory retrieval, migrate the setup to another OpenClaw installation, or compare Memora with builtin memory search.

Memora for OpenClaw

Resolve skill_dir as the directory containing this SKILL.md. Run bundled scripts from that directory; do not assume the current working directory.

Install and activate

1. Run bash "$skill_dir/scripts/setup.sh" --check-only.

2. If Ollama is stopped, rerun with --start-ollama.

3. If bge-m3 is absent, rerun with --pull-model. Warn that the model is roughly 1.2 GB before downloading it.

4. For a complete requested installation, run:

   bash "$skill_dir/scripts/setup.sh" \
     --start-ollama \
     --pull-model \
     --apply-config \
     --sync
   

5. Explain that the first --sync invokes the configured GLM route to extract memories and may incur API usage. Continue automatically when the user explicitly asked for installation plus indexing.

6. Run bash "$skill_dir/scripts/verify.sh" before reporting completion.

The setup script backs up ~/.openclaw/openclaw.json before changing it. It uses the full active-memory policy when that plugin exists and otherwise applies a minimal configuration that exposes memora_recall.

Operate the index

Use the plugin-local virtual environment:

"$skill_dir/.venv/bin/python" "$skill_dir/adapter.py" status
"$skill_dir/.venv/bin/python" "$skill_dir/adapter.py" sync
"$skill_dir/.venv/bin/python" "$skill_dir/adapter.py" sync --force

• Use ordinary sync after Markdown memory changes; it skips unchanged input.
• Use --force only for an embedding-model change, damaged index, or explicit rebuild request.
• Treat MEMORY.md and memory/**/*.md as the source of truth. Never edit or delete them while maintaining the Memora index.

For a private smoke test, run one semantic query and inspect only whether the expected fact is present:

"$skill_dir/.venv/bin/python" "$skill_dir/adapter.py" query \
  "<user-approved memory question>" --strategy semantic --top-k 5

Do not paste raw private memories into public logs, issues, or repositories.

Diagnose

Run:

bash "$skill_dir/scripts/verify.sh"
openclaw memory status --json

Interpret failures in this order:

1. Ollama server unavailable or bge-m3 missing.
2. Plugin-local .venv or pinned vendor/Memora missing.
3. OpenClaw plugin not loaded or configuration invalid.
4. Index missing or stale.
5. GLM route unavailable during synchronization.

Do not rebuild until the failing prerequisite is confirmed.

Compare with builtin memory

Use identical questions and top-k for both systems:

openclaw memory search --agent main --json --max-results 5 "<question>"
"$skill_dir/.venv/bin/python" "$skill_dir/adapter.py" query \
  "<question>" --strategy semantic --top-k 5

Measure evidence completeness, first relevant rank, and wall latency. For final answer tests, force memory_search or memora_recall separately and require the answer to be supported by that tool's returned evidence. Keep any personal-corpus evaluation set outside this repository.

Roll back

Prefer reversible disablement:

openclaw plugins disable memora-openclaw
openclaw gateway restart

Restore the timestamped configuration backup created under ~/.openclaw/backups/memora-openclaw/ only when the user explicitly requests restoration. Do not delete ~/.openclaw/memora-openclaw unless explicitly asked; it contains only the derived index and can be rebuilt.