Memory Orchestrator

Memory Orchestrator

Use this skill when the answer depends on ClawXMemory rather than only the current chat turn.

Primary Path

1. If the user is asking what memory exists or what ClawXMemory currently remembers, use memory_overview or memory_list.
2. Otherwise call memory_search with a concise query.
3. Read the returned:
   • route
   • context
   • selectedProjectId
   • disambiguationRequired
   • refs.files
   • debug
4. If context already contains the needed evidence, answer from it directly.
5. If exact verification is still useful, call memory_get with the smallest possible subset of refs.files.

Current Tool Contract

memory_search(...) returns the live retrieval contract:

• route
  • none
  • user
  • project_memory
• context
  • The rendered ClawXMemory recall block for this turn
• selectedProjectId
  • The formal project chosen for recall, or null
• disambiguationRequired
  • true when the query looks project-oriented but no formal project was selected
• refs.files
  • Relative file ids that can be passed into memory_get
• debug
  • Trace-oriented diagnostic metadata

Do not assume any old l2 / l1 / l0 / enoughAt fields exist.

Clarification Guardrail

If memory_search says:

• disambiguationRequired = true, or
• route = project_memory but selectedProjectId = null, or
• refs.files is empty for a project-specific question

then you must:

1. Say that no single formal project was selected from memory.
2. Ask a focused clarification question.
3. Do not list or guess project names that are not explicitly present in retrieved memory.

Example clarification:

• "我还没有选中唯一的项目。你说的是哪个项目？"
• "这个问题需要先确认具体项目名，你是在说哪个项目？"

Tool Usage Notes

• Prefer concise query strings.
• Keep memory_search.limit small unless the user explicitly wants breadth.
• Use memory_get only with ids returned from memory_search or memory_list.
• If evidence is missing, say so directly instead of filling gaps from memory guesses.
• When retrieved evidence conflicts, prefer the newest retrieved file and mention the conflict.

Example

User asks: "这个项目现在的阶段和下一步是什么？"
1) memory_search({ query: "这个项目现在的阶段和下一步是什么？", limit: 5 })
2) If disambiguationRequired=true, ask which project the user means.
3) Otherwise answer from context.
4) Only if needed, verify one or two refs.files with memory_get.

Guardrails

• Do not fabricate details not supported by retrieved memory.
• Do not fabricate or enumerate project names when retrieval did not select one.
• Do not call memory_get with guessed ids.
• If retrieval is empty or unresolved, ask a targeted follow-up instead of pretending memory certainty.