---

name: memory-palace-openclaw description: Use the Memory Palace OpenClaw plugin when the task is about durable memory recall, explicit memory verification, visual memory storage, plugin health/index maintenance, or chat-guided onboarding/bootstrap/provider configuration inside OpenClaw. metadata: { "openclaw": { "emoji": "🧠" } }

Memory Palace OpenClaw

Use this skill inside OpenClaw when the active memory plugin is memory-palace.

Default behavior

On a hook-capable OpenClaw host, the default path is:

• the active memory plugin is already memory-palace
• before_prompt_build is the primary automatic recall hook
• older hosts fall back to before_agent_start
• durable / reflection recall now search with the current session included, so recent in-thread context is less likely to be dropped
• a subset of user messages may be auto-captured after a successful turn
• visual context may be auto-harvested during the current turn

Current host boundary:

• automatic recall / capture / visual-harvest require typed lifecycle hooks from the host
• the current supported floor is OpenClaw >= 2026.3.2
• if the host lacks those hooks, treat openclaw memory-palace ... as the primary path and do not assume the default auto path is active

So the normal starting point is:

• trust the default recall / capture path first
• only intervene explicitly when the user asks for it, or when the default path is clearly not enough
• do not upgrade ordinary day-to-day requests into explicit tool calls unless there is a clear reason

Important boundary:

• text durable memory may be auto-recalled and auto-captured
• visual context may be auto-harvested
• visual memory is not auto-stored as a long-term record unless you explicitly use memory_store_visual
• the stable user-facing entry is the memory slot plus openclaw memory-palace ..., not raw MCP tools
• if command:new reflection or smart extraction cannot identify the target session transcript, the current behavior is to skip that transcript fallback instead of scanning the latest unrelated transcript
• compact-context recall is gist-only on the visible prompt path; do not surface stored metadata such as session_id, source_hash, or gist_method in chat replies
• lifecycle hook failures are swallowed and logged; after 5 consecutive failures for the same hook, that hook cools down for 60 seconds before one retry

Onboarding boundary:

• if the user wants to install, bootstrap, reconfigure, or verify provider readiness through chat, prefer the onboarding tools first
• do not send the user to the dashboard by default when the request is explicitly chat-first or VPS/terminal-only
• when Profile C/D is requested, provider probe comes before apply
• if provider probe fails or required fields are missing, explain the exact gap and fallback behavior in plain Chinese before suggesting apply

Advanced features (Sprint 1-3)

Retrieval and storage capabilities beyond the default path:

• RRF hybrid fusion — set RETRIEVAL_FUSION_METHOD=rrf to enable Reciprocal Rank Fusion in place of the default weighted-sum fusion. Treat it as opt-in: the May 15, 2026 rerun showed real C/D gains in memory-native and some retrieval metrics, but B and individual metrics were not uniformly better.
• Bi-temporal validity — every memory now carries valid_from / valid_until. Expired facts are excluded from search by default; pass include_expired=true only when the user explicitly wants historical / superseded facts.
• Guard IGNORE action — write guard actions are now five-valued: ADD / UPDATE / NOOP / DELETE / IGNORE. IGNORE means the LLM judged the input not worth storing (distinct from NOOP, which means "already stored and no update needed"). Surface the guard reason verbatim; do not retry with force=true unless the user re-confirms.
• Lifecycle hooks via .claude/settings.json — Claude Code hosts can wire passive memory helpers without explicit manual tool calls:
  • SessionStart → auto-load system://boot
  • PostToolUse on Edit|Write → emit a safe capture hint for non-sensitive project file edits
  • Stop → auto-trigger compact_context_reflection When these hooks are active, treat boot loading, edit hints, and reflection as host-driven; do not duplicate them with explicit MCP calls.
• Typed exception system — backend errors now surface structured codes from mcp_errors.py and sqlite_errors.py. When a tool call fails, quote the error code (not the raw message) so the user can match it against the troubleshooting docs.
• Memory bidirectional links — memories may carry typed links (related / supersedes / derived_from / contradicts). When recalling a fact, also surface its supersedes / contradicts targets so the user sees the resolution chain, not an isolated snippet.
• Provenance tracking (schema ready, partial wiring) — memory schema supports created_by_agent, created_by_session, and source_operation fields. Currently source_operation is populated on create_memory; agent/session fields require upstream host integration. When provenance fields are present, surface them; when absent, report as "not yet tracked".

When to intervene explicitly

Use explicit tools only in these cases:

1. The user clearly asks to check whether something is already in memory.
2. The user wants the exact content of a remembered item.
3. The user explicitly asks you to remember a stable fact, preference, or workflow for future turns.
4. The user wants to persist an image-derived summary as a long-term visual memory.
5. The user asks to check backend health, diagnose degradation, or rebuild the index.
6. The user asks to install/bootstrap/reconfigure Memory Palace through chat without opening the dashboard.

Preferred intervention order

1. Prefer the default OpenClaw memory path first.
2. If the user explicitly says “remember this” / “记住这个” and the content is a stable long-term fact, use memory_learn.
   • If memory_learn succeeds, acknowledge it with a minimal confirmation like 记住了。 / Stored. and move on.
   • If the user explicitly asked for an exact confirmation phrase after storing, use that phrase verbatim instead of paraphrasing it.
   • If memory_learn returns a blocked write, explain the brief human-readable reason.
   • Only if the user confirms they still want a separate durable memory saved should you rerun memory_learn with force=true.
3. If you need explicit recall, use memory_search.
4. If a hit matters, follow with memory_get using the returned path.
5. If the user wants a long-term visual record, use memory_store_visual.
6. If the user wants onboarding/bootstrap through chat, use:
   • memory_onboarding_status
   • memory_onboarding_probe
   • memory_onboarding_apply
7. If the problem is operational, ask the operator to run:
   • openclaw memory-palace status
   • openclaw memory-palace verify
   • openclaw memory-palace doctor
   • openclaw memory-palace smoke
   • openclaw memory-palace index

Use these explicit intervention tools:

• memory_search
• memory_learn
• memory_get
• memory_store_visual
• memory_onboarding_status
• memory_onboarding_probe
• memory_onboarding_apply

Anti-patterns

• Do not treat every request as a manual memory workflow.
• Do not treat this OpenClaw plugin skill as a mirror of the canonical MCP skill.
• Do not start with memory_search when the default recall path is already enough.
• Do not rely only on implicit auto-capture when the user explicitly asks you to remember a stable fact for future turns; use memory_learn.
• Do not use force=true on memory_learn unless the user has already confirmed they still want the memory saved after a blocked write.
• Do not invent memory contents.
• Treat backend/tool errors as real failures, not as empty recall.
• Do not imply that visual auto-harvest means long-term visual auto-storage.
• Do not present raw MCP tools as the default entry for ordinary OpenClaw users.
• Prefer text summaries and references for visual memory, never raw binary image storage.
• If you explicitly recall something important, prefer memory_search first and memory_get second.
• Do not tell VPS/terminal-only users that dashboard setup is required when onboarding tools can do the same job.
• Do not claim Profile C/D is ready before provider probe actually passes.
• Do not hide fallback to Profile B; name the failing provider and tell the user what to fix.
• Do not tell users that /responses is the primary LLM runtime path; current main path is still /chat/completions.
• Do not switch fusion to RRF silently; only set RETRIEVAL_FUSION_METHOD=rrf when the user asks or when default fusion is clearly under-performing.
• Do not include expired facts in recall by default; flip include_expired=true only on explicit user request for historical context.
• Do not conflate IGNORE with NOOP; IGNORE is a deliberate "not worth storing" decision, NOOP is "already stored".
• Do not duplicate host-driven lifecycle hooks (SessionStart / PostToolUse / Stop) with manual boot/edit/reflection calls.
• Do not invent provenance, links, or temporal fields when they are absent; report them as missing instead.
• Do not leak compact-context metadata into visible chat; only use the gist text that is meant for recall.

Trigger examples

Default recall path:

• “回忆一下之前记过的偏好”
• “继续我们上次的决定”

Explicit verification:

• “查记忆里有没有这件事”
• “把刚才提到的那条记忆正文读出来”

Explicit durable remember:

• “请记住这个长期偏好：以后默认简洁回答”
• “Remember this workflow for future turns: code first, tests immediately after, docs last”
• “请记住这个长期偏好，存好后只回复‘已记录。’”
• “Remember this durable fact and reply only ‘stored profile c’ after saving”

Exact confirmation phrase:

• If the user explicitly requires one exact confirmation string after memory_learn, pass it through confirmationPhrase (or confirmation_phrase) and reuse that phrase verbatim.
• Example: memory_learn({ content: \"请记住这个长期偏好：以后默认简洁回答\", confirmationPhrase: \"已记录。\" })
• Example: memory_learn({ content: \"Remember this workflow for future turns\", confirmationPhrase: \"stored profile c\" })

Long-term visual storage:

• “把这张白板照片的 OCR 摘要存成 visual memory”
• “把这张会议截图整理成长期 visual record”

Maintenance and diagnostics:

• “检查 memory backend 状态并重建索引”
• “跑一遍 verify / doctor / smoke 看插件是不是正常”

Chat-first onboarding:

• “不要开 dashboard，直接通过对话把 Memory Palace 接进 OpenClaw”
• “先帮我判断该走 Profile B 还是 C/D，再一步一步收集 provider 配置”
• “我在 VPS 上，只能通过聊天引导完成 setup”

Advanced features:

• “切换到 RRF 融合，关键词和语义结果差太多”
• “这条事实已经过期了吗？把历史版本也调出来看看” (include_expired=true)
• “这条信息为什么没存？是 IGNORE 还是 NOOP？”
• “这条记忆是谁、什么时候、哪个 session 写进去的？” (provenance)
• “找出和这条记忆 contradicts / supersedes 的其他记忆” (bidirectional links)
• “Claude Code 的 SessionStart / PostToolUse / Stop 钩子是不是已经接上了？”