Feishu Group Memory

Architecture

Scripts handle data only — no LLM calls inside scripts:

• - onboarding.py — read/write industry knowledge pack config
• INLINECODE1 — fetch raw messages from Feishu; save analyzed records
• INLINECODE2 — keyword search over stored records
• INLINECODE3 — SkillPay charge/balance/payment-link

All AI analysis is done by you (the OpenClaw model): deciding what to record, extracting structured fields, generating advice, writing summaries.

Quick Reference

First Use: Onboarding

At the start of every session, check whether an industry pack is configured:

CODEBLOCK0

• - {"configured": true, "context": "..."} → load the context field and proceed
• INLINECODE14 → run onboarding before anything else

Onboarding conversation

Ask the user:

"Before we start, I'd like to understand what your group is mainly used for so I can record and analyze the right things.
Choose one, or describe it in your own words:

1. 1. 📈 Sales tracking (leads, quotes, deals)
2. 🎧 Customer support (tickets, issues, complaints)
3. ⚖️ Legal matters (contracts, risks, cases)
4. 📋 Project management (tasks, milestones, blockers)
5. ✍️ Describe my own use case"

Saving the config

Built-in industry (options 1–4):

python3 {baseDir}/scripts/onboarding.py setup \
  --industry sales \
  --workspace ~/.openclaw/workspace

Custom description (option 5):

Using the user's description and the template at {baseDir}/templates/context-template.md, generate the knowledge pack yourself, then save it:
CODEBLOCK2

Confirm with the user: "Got it! I'll use this context going forward. Which group would you like me to start recording?"

Feature: Record Group Messages

Trigger: "record X group", "fetch messages from X", "capture what's been discussed in X"

Step 1 — Find the group

python3 {baseDir}/scripts/listener.py find_chat --name "KEYWORD"

Step 2 — Fetch raw messages

python3 {baseDir}/scripts/listener.py fetch_raw \
  --chat_id CHAT_ID \
  --limit 100 \
  --workspace ~/.openclaw/workspace

Step 3 — You analyze

• - Is it worth recording?
• What category does it belong to?
• Who or what is the key entity (person, company, project)?
• What structured fields can be extracted?
• What is the urgency (high / medium / low)?

Step 4 — Save the records

Step 5 — Report to user

"Analyzed 100 messages. Saved 12 items:

• - 3 customer intent signals (Li, Wang, Chen)
• 5 follow-up actions
• 4 pricing discussions

2 high-urgency items — want me to walk through them?"

Billing

python3 {baseDir}/scripts/billing.py charge \
  --user_id USER_ID --amount 0.005 --label "message analysis"

Feature: Query Records

Trigger: "how is Wang doing", "what happened with Acme last week", "show me recent follow-ups"

CODEBLOCK7

Returns matching records as raw JSON. You turn them into a natural-language answer, e.g.:

"Here's what I have on Wang (Wang Zong):

• - Jan 15: Said he can sign next week (high priority)
• Jan 12: Asked about discount options, still considering

Last contact was 3 days ago — worth reaching out today."

No charge for queries.

Feature: AI Advice

Trigger: "how should I follow up with X", "give me some advice", "help me think through this"

First, search for relevant records:
CODEBLOCK8

Then reload the industry pack if needed:
CODEBLOCK9

Using the "Advice Templates" section of the knowledge pack and the retrieved records, give the user concrete, actionable advice directly.

CODEBLOCK10

Feature: Summary Report

Trigger: "summarize today", "weekly report", "what happened this week"

CODEBLOCK11

You write the summary. Example structure:

Weekly Summary (Jan 13–19)
28 items recorded across 7 customers.
Action required (3)

• - Li Zong: ready to sign — prepare draft contract
• Wang Zong: price sticking point — request special approval

By category

• - Customer intent: 12 | Follow-ups: 8 | Pricing: 5 | Other: 3

Suggestion: 2 customers haven't been contacted in 5+ days.

CODEBLOCK12

Error Handling