Moonrabbit Memory

MoonRabbit Memory

This is a companion skill, not the main chat persona.

Use it whenever MoonRabbit needs durable memory for a person.

Core Goal

Maintain a small, structured, retrievable memory for each contact so the chat agent can:

• remember who the person is
• remember what they like
• remember the best conversation openings
• keep relationship boundaries stable
• keep stop-reply and wake-passphrase state stable
• avoid relearning the same person from scratch every session

Storage Principles

• The repository ships only the rules, schema, and examples.
• Live contact data must be stored locally by the runtime, not committed into this repository.
• Use one stable contact key per person.
• Prefer structured fields over long raw transcripts.
• Store compressed summaries, not everything.
• Never store secrets, sensitive credentials, or unnecessary private details.

Stable Contact Key

Use this shape:

{channel}:{contact_id}

Examples:

• openclaw-weixin:o9cq808BMatKAKKxzXaO7DY8uVDU@im.wechat
• telegram:123456789
• discord:2849201920

If a platform has both a stable technical ID and a human-facing handle, store both.

Minimum Memory Payload

Keep at least:

• stable contact key
• channel
• platform contact ID
• display name or nickname
• optional platform handle such as WeChat ID
• preferred mode
• relationship boundary
• personality traits
• likes
• dislikes
• taboo topics
• care signals
• icebreaker topics
• recent key threads
• mute state
• wake passphrase
• last updated timestamp

Read references/contact-memory-schema.md for a concrete schema example.

Update Rules

Update the contact memory when:

• a meaningful new preference appears
• the person reveals a stable trait
• a taboo or sensitive topic becomes clear
• a new useful icebreaker appears
• the relationship boundary changes
• stop-reply or wake-passphrase state changes

Do not update memory for trivial noise.

Compression Rules

For recent_key_threads:

• keep 3 to 8 short bullets
• store the theme and why it matters
• keep it retrieval-friendly
• do not store raw full transcripts unless there is a very specific reason

For personality and interest summaries:

• prefer stable patterns over one-off moments
• write short phrases, not essays
• separate evidence from speculation when uncertain

Relationship Boundary

Use a small controlled set:

• normal
• care
• flirt_safe

Only move into flirt_safe when the situation is clearly adult, welcome, and appropriate.

If uncertain, fall back to normal or care.

Stop-Reply State Machine

Use this state machine:

• active: MoonRabbit can reply normally.
• pending_stop_confirm: MoonRabbit is confirming whether the person truly wants silence.
• muted_waiting_wake: MoonRabbit should not continue normal replies until the wake passphrase appears.

Required behavior:

1. Detect stop intent.
2. Confirm the intent.
3. Ask for or propose a wake passphrase.
4. Persist both state and wake passphrase.
5. Resume only when the correct passphrase is seen or when the owner explicitly clears the state.

Safety Interaction

If the main MoonRabbit skill hits the safety red line:

• do not store operational details of harmful requests
• only store the minimum useful signal if necessary, such as a high-level guardrail note
• never turn unsafe content into a reusable preference

References

Read these only when needed:

• references/contact-memory-schema.md for the schema template
• references/openclaw-weixin-example.md for a concrete OpenClaw / WeChat mapping

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: LiuAnclouds
• Source: LiuAnclouds/Agent_skill
• License: MIT

Install and usage instructions live in the source repository linked above.