agi-is-going-to-arrive/memory-palace-openclaw/memory-palace-openclaw-onboarding

Memory Palace OpenClaw Onboarding

Use this skill when the user wants to install, bootstrap, reconfigure, or troubleshoot Memory Palace onboarding through normal OpenClaw conversation.

This skill is only for onboarding.

• It is not the day-to-day durable memory skill.
• It should not replace memory-palace-openclaw.
• It should assume the user may be on a VPS with no dashboard access.

Default posture

• Prefer tool-driven conversational onboarding first.
• Use the onboarding tools before falling back to raw CLI commands.
• Do not require the dashboard unless the user explicitly asks for it.
• Prefer stdio transport unless the user already knows they need sse.
• Prefer Profile C/D when the user can provide real providers.
• Fall back to Profile B only when providers are missing, failing, or the user explicitly wants the safest bootstrap path.

Primary onboarding tools:

• memory_onboarding_status
• memory_onboarding_probe
• memory_onboarding_apply

Locale rule:

• if the user is speaking Chinese, call these tools with locale="zh-CN"
• if the user is speaking English, call these tools with locale="en"

Supported provider input forms

When collecting provider values, accept these forms:

Embedding

• Base URL:
• https://provider.example/v1
• https://provider.example/v1/embeddings
• API key:
• any normal bearer-style secret string
• Model:
• exact provider model id, for example Qwen3-Embedding-8B

Reranker

• Base URL:
• https://provider.example/v1
• https://provider.example/v1/rerank
• API key:
• any normal bearer-style secret string
• Model:
• exact provider model id, for example Qwen/Qwen3-Reranker-8B

LLM

• Base URL:
• https://provider.example/v1
• https://provider.example/v1/chat/completions
• https://provider.example/v1/responses
• API key:
• any normal bearer-style secret string
• Model:
• exact provider model id, for example gpt-5.4

Interpretation boundary:

• OpenAI-compatible chat-style configuration is supported.
• If the user pastes a full /chat/completions or /responses URL, treat it as valid input and normalize it to the base automatically.
• The current runtime accepts /responses as an input alias, but the main write-guard / compact-gist LLM flows still call /chat/completions.

What to ask, in order

Ask only for what is still missing.

1. Whether they want the safest bootstrap path or the strongest capability path.
2. Whether they need basic or full.
3. Whether they can provide real embedding, reranker, and LLM providers right now.
4. The missing provider fields section by section.

Keep the wording simple:

• “If you can provide real providers now, I recommend Profile C or D.”
• “If not, we can start with Profile B first and upgrade later.”

Preferred execution flow

1. Start with memory_onboarding_status.
2. If the user wants the strongest path and has provider values, run memory_onboarding_probe.
3. Read the probe result before applying setup.
4. If the embedding probe reports recommendedDim, detectedMaxDim, or detectedDim, explicitly tell the user:

• what dimension was detected
• that this is the recommended RETRIEVAL_EMBEDDING_DIM

1. Apply setup only after the probe result is understood:

• memory_onboarding_apply

1. If onboarding tools are unavailable, fall back to raw CLI:

• python3 scripts/openclaw_memory_palace.py onboarding --json
• python3 scripts/openclaw_memory_palace.py provider-probe --profile c|d ... --json
• python3 scripts/openclaw_memory_palace.py setup ... --json
• On Windows PowerShell, replace python3 with py -3

1. After setup, always verify:

• openclaw memory-palace verify --json
• openclaw memory-palace doctor --json
• openclaw memory-palace smoke --json

How to interpret probe/setup results

If provider probe passes

• Tell the user C/D is ready to apply.
• If embedding dimension was detected, recommend using that value.

If provider probe reports missing fields

• Tell the user exactly which fields are still missing.
• Ask only for those fields.
• Do not claim that C/D is ready.

If provider probe reports failed providers

• Tell the user which provider failed and why.
• Ask them to fix the address, key, model, or reachability.
• Re-run provider-probe before setup.

If setup returns fallbackApplied=true

• State clearly that the requested C/D profile did not stay active.
• State that the environment is now on Profile B.
• Explain that this is a temporary safe fallback, not the target capability path.
• Tell the user to re-probe and then re-run setup after fixing providers.

Profile guidance

Preferred path

Strongly prefer Profile C/D when the user can provide real providers.

Why:

• real embedding
• real reranker
• better retrieval quality
• stronger end-state capability

Safe bootstrap path

Profile B is still the safe bootstrap baseline.

Its hard retrieval boundary is:

• retrieval stays on local hash embeddings
• embedding dimension stays on the local baseline
• reranker stays off

But Profile B can still keep optional LLM-assisted features when valid LLM config is present.

That means:

• write guard can still be enabled
• compact gist can still be enabled
• retrieval quality still does not become true provider-backed C/D retrieval

If no LLM config is provided, those LLM-assisted features stay off and the system continues on the non-LLM path.

Interaction rules

• Do not force the user into the dashboard.
• Do not ask for all fields at once if the probe already narrowed the missing set.
• Do not invent provider values.
• Do not echo secrets back unless the user explicitly asks for the literal command line.
• Prefer the onboarding tools when they are available.
• If tool execution is unavailable, provide the exact commands the user should run.
• If shell execution is available but onboarding tools are also available, still prefer the tools so the conversation stays simple for VPS users.

Anti-patterns

• Do not treat the runtime memory skill as the onboarding skill.
• Do not say C/D is ready before running provider-probe or equivalent validation.
• Do not hide fallback to B.
• Do not describe Profile B as equivalent to C/D.
• Do not assume the user can open /setup or any graphical page.

Common Errors

session file locked

If the user runs a local CLI agent turn and hits:

session file locked

Explain it directly:

• another OpenClaw process is already writing the same session file
• the usual cause is an already-open WebUI or another CLI/gateway process using that same agent/session

Preferred recovery:

1. use a fresh session or a temporary agent for the test turn
2. or close the process currently holding that session lock

HTTP 401 ... authentication token has been invalidated

If the conversational turn fails with model-auth 401, do not describe it as the onboarding skill failing.

Explain:

• the onboarding skill and onboarding tools were already loaded
• the failure is in the host's current chat-model provider authentication

Preferred recovery:

1. refresh or re-login the current model provider token
2. or switch the agent to a currently working model, then retry the onboarding conversation

provider probe failure

If a provider probe returns fail, always tell the user:

1. which provider failed
2. whether the likely cause is base URL, API key, model name, or reachability
3. that default apply will only fall back to Profile B temporarily
4. that this does not mean Profile C/D is actually ready

Trigger examples

• “帮我把 memory-palace 接进 OpenClaw，不开 dashboard。”
• “我在 VPS 上，只能对话配置，你来引导我装。”
• “帮我配置 Profile C。”
• “provider 报错了，先看哪里没填好。”
• “先帮我探测 embedding 维度，再决定怎么填 env。”