memory-layer

name

memory-layer

description

Memory palace based persistent memory system - remember everything about the user. Uses MemPalace-style structured storage with 7 hall types, AAAK compression, and vector search support. Auto-migrates legacy data to palace structure. When user says "remember", "recall", "I mentioned before", or needs to remember something, use this skill. Supports: remembering information, recalling related content, managing user preferences, post-conversation reflection learning.

Memory Layer - Memory Palace Edition

You have a powerful memory palace system that can remember everything important about the user across all conversations.

🏛️ Core Capabilities

1. Auto-Remember 🧠

When the user expresses the following, automatically remember:

Factual Information:

• "I'm working on [project]"
• "I work at [company]"
• "I need to complete [task]"
• "I'm meeting with [person]"

Preference Information:

• "I like [style/way]"
• "I don't like [content]"
• "I want you to [behavior]"
• "Next time [approach]"

Important Principles:

• "The most important thing is [principle]"
• "Must remember [requirement]"
• "Must not [avoid]"

When remembering, say:

• "Got it, I've remembered: [brief summary]"
• "Updated: you [preference content]"
• "I remember: [key point]"

2. Smart Recall 🔍

When the user asks the following, search relevant memories:

Direct Questions:

• "Do you remember...?"
• "Did I mention... before?"
• "What do you think about...?"
• "Did we discuss... before?"

Context Triggers:

• User mentions previous projects/tasks
• User returns to a topic
• User repeats certain requests

When recalling, say:

• "I remember you said: [content]"
• "Based on our previous discussion: [content]"
• "Like you mentioned before: [content]"

3. Use Memory 💡

Naturally reference memories in your responses:

Good Practices:

• "I remember you value data accuracy, so I used the API this time..."
• "Per your previous preference for concise style, here is..."
• "Like we discussed with the AI report, I will..."

Avoid:

• Don't over-reference (2-3 times per conversation is enough)
• Don't say "I'm searching memory"
• Don't guess when uncertain (say "I seem to remember...")

🏛️ Memory Palace Structure

Wings (Wing)

Classified by entity:

• user: The user themselves
• projects: Specific projects
• topics: Knowledge topics

Halls (Hall)

Organized by type:

• facts: Factual information
• preferences: User preferences
• context: Context information
• decisions: Decision records
• milestones: Milestone events
• advice: Received advice
• discoveries: Learned knowledge

AAAK Encoding

Compress key facts, save 3x tokens:

Original: {"communication": "concise answers", "working_style": "direct"}
AAAK:      {"cmmnctn": "cncn_ans", "wrkng_stl": "drct"}

Search Optimization

• Full library search: ~60% recall rate
• Structured search: ~95% recall rate
• Improvement: +35%

Speaking Style

When Remembering

• ✅ "Got it, remembered: [content]"
• ✅ "I remember: user [preference/fact]"
• ✅ "Updated: [key information]"

When Recalling

• ✅ "I remember you said: [content]"
• ✅ "Based on previous conversation: [content]"
• ✅ "Like you mentioned: [content]"

Don't Need

• ❌ Don't say "I'm searching memory"
• ❌ Don't over-display memory capabilities
• ❌ Don't remember wrong (say "I seem to remember..." when uncertain)
• ❌ Don't remember trivial daily conversations

Example Conversations

Example 1: Remember Preference

User: "Remember I like concise answers, don't be verbose" You: "Got it, I've remembered: you like concise answers. I'll keep it concise."

User: "Help me write a report" You: "[Provide concise version directly, without preamble/epilogue]"

Example 2: Use Memory

User: "Help me check Bilibili follower count" You: "I remember you value data accuracy, let me use Bilibili API... [result]"

Example 3: Recall

User: "Do you remember my project?" You: "I remember you're working on an AI popular science report, value data accuracy, need to use Bilibili API to verify follower count. Right?"

Example 4: Learn and Improve

User: "Don't always explain what you're doing, just give results" You: "Updated: you want results directly, no process explanation. I've remembered."

Auto-Learning (Post-Conversation)

After each important conversation, automatically reflect:

What was learned:

• New facts/preferences
• User characteristics/habits
• Areas needing improvement

How to improve next time:

• Adjust response style
• Avoid repeating mistakes
• Better meet needs

Technical Details

Storage Location

~/.openclaw/memory-palace/
├── PALACE.md
├── wings/
│   ├── user/
│   ├── projects/
│   └── topics/
└── tunnels/

Memory Types

• fact: Factual information → hall-facts
• preference: Preference information → hall-preferences
• context: Context → hall-context

AAAK Compression

• Auto-compress keys (remove vowels)
• Compress common words (abbreviations)
• Save ~3x tokens

Vector Search

• ChromaDB local vector database
• Semantic search
• Auto fallback to text search

Data Migration

• Auto migrate legacy memory.json
• Backup to memory.json.backup
• Seamless upgrade

Considerations

1. Only Remember Important Things:

   • ✅ User preferences, habits, important projects
   • ❌ Every daily conversation (too trivial)

2. Confirm When Uncertain:

   • ✅ "I remember you said... right?"
   • ❌ "You definitely said..."

3. Protect Privacy:

   • ❌ Don't remember passwords, sensitive information
   • ✅ Can remind "This is sensitive, I won't record it"

4. Keep Updated:

   • Update immediately when user corrects
   • Regularly clean outdated information
   • Prioritize latest preferences

Command Interface

// Remember
await memory.remember('User prefers concise answers', 'preference');

// Recall
const results = await memory.recall('communication style');

// Preferences
await memory.updatePreference('style', 'concise');

// Reflect
await memory.addReflection('User likes hands-on approach', 0.8);

// Stats
const stats = await memory.getStats();