stack/skills/memory-discipline/SKILL.md

name, description, metadata

name	description	metadata
memory-discipline	Enforce proper session memory recording. Use at: (1) session idle timeout, (2) task completion, (3) explicit /save command, (4) end of significant conversation milestones. Extracts decisions, key info, and action items from the session and writes them to memory storage (files or database).	openclaw emoji requires 💾

Memory Discipline Skill

This skill ensures that session knowledge is properly captured and stored, not lost. It is mandatory to run this before:

• Session idle timeout (no user input for X minutes)
• Task completion
• Explicit /save or /memory command
• Any significant milestone in the conversation

Why This Exists

Too often agents don't record important information from sessions. This causes:

• Repeated questions about things the human already told you
• Lost context between sessions
• Frustration when "I told you that last time" happens

This skill makes memory recording REQUIRED, not optional.

Workflow

Step 1: Extract Key Information

Before writing, analyze the session and extract:

Category	What to capture
Decisions	Any decision made by the human or jointly (e.g., "we'll use Postgres + pgvector")
Important Info	Facts the human provided that should be remembered (e.g., preferences, constraints)
Action Items	Things to do later, follow-ups, tasks not yet completed
Context Changes	Project state changes, new businesses, new channels
Lessons	What worked, what didn't, what to avoid

Step 2: Write to Storage

File-based (current):

• Update memory/YYYY-MM-DD.md with new entries
• Update MEMORY.md for long-term memory
• Update USER.md if human preferences changed
• Update project-specific files if applicable

Database (future - when Postgres/pgvector available):

• Write to session_summaries table
• Generate embeddings for RAG context
• Log to event_log table

Step 3: Confirm in Chat

After saving, say EXACTLY:

💾 Saved to memory:
- X decisions
- X key info items
- X action items

[If action items exist:]
TODO: [list each action item]

This confirms to the human that their information was saved.

Triggers

Automatic triggers (run yourself):

1. Session idle for 5+ minutes (check on heartbeat or cron)
2. Task marked as complete
3. Human explicitly asks to save

Manual triggers (human invokes):

• /save - Save current session state
• /memory - Show what has been saved this session
• /remember <info> - Manually add something to memory

Quality Standards

✅ DO:

• Write specific, actionable entries
• Include context (why this matters)
• Use bullet points for readability
• Include dates when relevant
• Add to appropriate memory file (daily vs long-term)

❌ DON'T:

• Write vague entries like "discussed project"
• Skip this step "if nothing important happened"
• Save trivial things (greetings, small talk)
• Write excessively long entries

Integration Points

When running this skill, also check:

1. TODOs - Any pending tasks that should be logged?
2. Open Threads - Any ongoing conversations that need continuation notes?
3. Project State - Did any project context change?

Example Output

💾 Saved to memory:
- Decision: Use Valkey (not RabbitMQ) for message bus
- Decision: bge-m3 for embeddings, cogito:14b for local chat
- Key info: Postgres image already has pgvector built-in
- Action: Check deployed stack via Portainer at 10.1.1.45

TODO:
- [ ] Pull bge-m3 model to Ollama
- [ ] Verify pgvector in postgres image
- [ ] Draft instance configs for Main + 3 departments