OpenClaw SiliconFlow Memory

Goal

Configure OpenClaw memory_search to use SiliconFlow embeddings reliably, then verify indexing and retrieval.

Use this skill to:

• - switch memory embeddings to SiliconFlow
• standardize on INLINECODE1
• set fallback: "none" for clearer debugging
• add curated external Markdown paths to semantic memory
• diagnose failed indexing or failed retrieval

Workflow

1. Inspect the memory config surface first

Before changing config, inspect the relevant schema paths:

• - INLINECODE3
• INLINECODE4
• INLINECODE5

Then inspect the current config with gateway config.get so the patch is based on the latest hash.

Important:

• - Configure memory under agents.defaults.memorySearch, not top-level memorySearch.
• For OpenAI-compatible embedding endpoints, set remote.apiKey explicitly.

2. Apply the recommended memorySearch config

Use gateway config.patch and keep the configuration explicit.

Recommended baseline:

CODEBLOCK0

Why this baseline works well:

• - provider: "openai" matches SiliconFlow's OpenAI-compatible embeddings API
• INLINECODE12 is a strong default for mixed Chinese/English Markdown recall
• INLINECODE13 avoids silently switching providers during debugging
• explicit remote.baseUrl and remote.apiKey remove ambiguity

3. Add external knowledge paths conservatively

When indexing a local knowledge base, prefer curated paths instead of adding the entire repo at once.

Good first candidates:

• - source notes
• concepts
• project overviews
• indexes
• README-style top-level knowledge summaries

Avoid adding these in the first pass unless the user explicitly wants them indexed:

• - inbox or capture folders
• daily journals
• templates
• archives
• noisy exports or generated files

Example:

CODEBLOCK1

4. Validate in this order

After config reload/restart, validate in a fixed order:

1. 1. INLINECODE16
2. INLINECODE17
3. INLINECODE18
4. tool-layer memory_search with the same query

Success signs:

• - INLINECODE20
• INLINECODE21 increases as expected
• INLINECODE22
• retrieval returns snippets from the intended external paths

5. Use known text for verification

Do not validate with vague keywords only.

Prefer a phrase that appears verbatim in the indexed notes, such as:

• - a README opening sentence
• a project title
• a section heading
• a unique domain phrase

This makes it easier to distinguish:

• - indexing failure
• retrieval quality issues
• query mismatch

Common failures

404 page not found

Interpret this as: the endpoint likely does not provide embeddings at the configured route.

Typical cause:

• - using a chat/completions proxy that does not expose INLINECODE24

Action:

• - switch to SiliconFlow's embeddings endpoint
• confirm remote.baseUrl is INLINECODE26
• confirm the embedding model name is valid

fetch failed

Interpret this as: the request could not complete at all.

Typical causes:

• - bad API key
• unreachable endpoint
• transient network issue
• provider auto-detection picked the wrong backend

Action:

• - make provider, model, remote.baseUrl, and remote.apiKey explicit
• retry INLINECODE32

EBUSY: resource busy or locked

Interpret this as: the SQLite memory store is locked during reindex.

Action order:

1. 1. retry the index once after current memory operations settle
2. restart OpenClaw and retry
3. if the lock persists, reboot the machine and retry

Do not assume config is wrong just because reindex failed with EBUSY.

Recommended operating style

• - keep the embedding provider explicit
• keep fallback: "none" during setup and debugging
• add knowledge paths in small batches
• validate with both CLI and tool-layer retrieval
• prefer high-signal Markdown sources over raw dumps

Completion checklist

Before declaring success, confirm all of the following:

• - SiliconFlow API key is configured
• INLINECODE36 is active
• INLINECODE37 shows INLINECODE38
• reindex succeeds
• indexed file count reflects the intended extra paths
• at least one query returns a result from the external knowledge base