RAG Memory Skill

Architecture (as of 2026-04-04)

All files live under ~/.openclaw/skills/rag-memory/ — the single source of truth.

Plugin is installed at ~/.openclaw/extensions/rag-memory/ and configured in ~/.openclaw/openclaw.json under plugins.entries.rag-memory.

Qdrant collections: {QDRANT_COLLECTION_PREFIX}_memory (daily logs + facts) and {QDRANT_COLLECTION_PREFIX}_docs (tools + skill docs). 768-dim, nomic-embed-text via OpenAI-compatible embedding endpoint.

When to use vector_search

Always prefer vector_search over reading full memory files:

• - Before reading any daily log or MEMORY.md → INLINECODE23
• Before reading TOOLS.md or any SKILL.md → INLINECODE24
• Unsure → INLINECODE25

Only fall back to reading the file directly if vector_search returns no results.

Sync management

Incremental sync also fires automatically when any .md file changes in ~/.openclaw/workspace/memory/ (via sysclaw-rag-watch.path). Nightly full sync runs at 03:00 (sysclaw-rag-sync.timer).

Updating the plugin

If plugin/index.js is changed:

CODEBLOCK0

If config.env tuning values change, also update the matching values in plugins.entries.rag-memory.config in openclaw.json (especially score_threshold and top_k).

Deploying systemd unit changes

Edit files in systemd/ (source of truth), then deploy using the included script — it substitutes the correct username automatically:

CODEBLOCK1

Or manually, replacing youruser with the local OpenClaw user:

CODEBLOCK2

Key config values

If Qdrant is unreachable

1. 1. Try once — do not retry in a loop
2. Fall back to reading the relevant file directly
3. Note in your response that vector search was unavailable
4. Check: INLINECODE54
5. Do not restart Qdrant automatically — alert the user

Not for

Do not activate this skill for:

• - General file search or code search unrelated to memory/RAG
• Editing files (use read/write tools directly)
• Tasks that don't involve looking up past context, facts, or tool docs
• Syncing or resyncing when the user has not asked to update memory
• Any request where the user explicitly says "don't search memory"

External endpoints

This skill sends data to three external services configured via config.env:

No data is sent to any Anthropic or third-party endpoint by this skill.

Privacy statement

Memory content (daily logs, facts, tool descriptions) leaves the local machine in two ways:

1. 1. Text is sent to your configured embedding endpoint (EMBED_BASE_URL) to generate vectors
2. Vectors + metadata are stored in and queried from your Qdrant instance (QDRANT_HOST)

Both services are self-hosted and configured by the user. No data is sent to any external party by this skill.

Security manifest

Environment variables accessed:

• - QDRANT_HOST, QDRANT_API_KEY, INLINECODE66
• INLINECODE67, EMBED_API_KEY, INLINECODE69
• INLINECODE70
• INLINECODE71, CHUNK_SIZE, CHUNK_OVERLAP, TOP_K_DEFAULT, SYNC_BATCH_SIZE, INLINECODE76

Files read:

• - All .md files under INLINECODE78
• INLINECODE79 (loaded at sync script startup)

Files written:

• - None (no local file writes)

Database writes (Postgres):

• - memory_entries.qdrant_synced_at — updated after successful vector upsert
• INLINECODE81 — same
• INLINECODE82 — same
• INLINECODE83 — one row inserted per sync run (recordssynced, durationms)

Network:

• - Outbound HTTPS to EMBED_BASE_URL (embedding)
• Outbound HTTPS to QDRANT_HOST (vector storage/search)
• Outbound TCP to Postgres host in INLINECODE86

Trust statement

By installing this skill, text from your memory files and database will be sent to the embedding endpoint configured in EMBED_BASE_URL. Vectors are stored in the Qdrant instance at QDRANT_HOST. Both are self-hosted services you configure. Only install if you trust those services with your memory contents.

When auto_inject is enabled (default: true), Qdrant search results are prepended to every prompt context. This data originates from your own self-hosted Qdrant instance — no external party supplies it. The content is sanitized of session boilerplate before storage but is otherwise injected as-is; only enable auto_inject if you trust your Qdrant data as prompt input.

License

MIT-0 — no attribution required.