recall-memory

Recall Memory

Retrieve information stored across sessions using memU's agentic memory framework.

Overview

This skill retrieves information from memU - a three-layer memory hierarchy that processes conversations into structured knowledge:

• Resources: Raw conversation data from past sessions
• Items: Extracted memory units (preferences, skills, opinions, habits, facts)
• Categories: Aggregated summaries with full traceability

How it works: Conversations are automatically memorized when sessions end (via SessionEnd hook). This skill retrieves that stored knowledge when needed.

Script Path Construction

IMPORTANT: Always use full paths to call scripts. Do NOT use cd to change to the scripts directory.

The skill is located at: Base directory for this skill (shown when skill loads)

To call scripts, concatenate:

• Skill base directory + /scripts/ + script name

Example:

# If skill base is: /home/user/.claude/plugins/memu/skills/recall-memory
# Then memu.py is at:
/home/user/.claude/plugins/memu/skills/recall-memory/scripts/memu.py

In examples below, we use scripts/memu.py as shorthand, but replace scripts/ with the full path based on the skill's base directory.

Retrieval Methods

memU supports two retrieval methods - choose based on query complexity:

RAG (Recommended)

When to use: Specific factual queries, direct questions, known topics

• Fast vector-based similarity search
• Returns results in <2 seconds
• Good for: "What are my preferences?", "Do I use TypeScript or JavaScript?", "How do I handle errors?"
• Best when you know what you're looking for

Example:

scripts/memu.py retrieve --query "What coding style do I prefer?" --method rag

LLM

When to use: Complex queries, ambiguous questions, need deep understanding

• Deep semantic understanding with adaptive refinement
• Slower (~5-10 seconds) but more thorough
• Good for: "What patterns do I follow when refactoring?", "How do I approach testing?", open-ended questions
• Best for nuanced or multi-faceted questions

Example:

scripts/memu.py retrieve --query "What patterns do I use in React components?" --method llm

Choosing: Start with RAG (faster). If results are insufficient or query is complex, use LLM.

The Process

Follow these steps to retrieve memory:

Step 1: Formulate Query

Convert user's question into a clear query string:

• User asks: "What do you know about my testing approach?"

• Query: "What is the user's testing approach?"

• User asks: "How do I usually name variables?"

• Query: "How does the user name variables?"

Step 2: Choose Method

• Simple/direct question → RAG
• Complex/ambiguous question → LLM

Step 3: Execute Retrieval

Call the script with appropriate method:

scripts/memu.py retrieve --query "your query here" --method rag

or

scripts/memu.py retrieve --query "your query here" --method llm

Step 4: Parse Response

The script returns JSON with this structure:

{
  "status": "ok",
  "data": {
    "categories": [
      {
        "name": "Coding Preferences",
        "summary": "User prefers TypeScript with strict typing..."
      }
    ],
    "items": [
      {
        "memory_type": "preference",
        "summary": "Prefers TypeScript over JavaScript"
      }
    ],
    "resources": [
      {
        "modality": "conversation",
        "url": "session-2024-01-15"
      }
    ]
  }
}

Step 5: Integrate Results

Extract relevant information from the response:

1. Check categories for high-level summaries
2. Check items for specific facts/preferences
3. Integrate into your response to the user

Command Reference

Command	Purpose	Speed
scripts/memu.py retrieve --query "..." --method rag	Fast retrieval (default)	<2s
scripts/memu.py retrieve --query "..." --method llm	Deep semantic retrieval	5-10s

Required parameters:

• --query: The question/search query (string)
• --method: Either rag or llm (default: rag)

Output: JSON to stdout with status and data fields

Output Formats

Success Response

{
  "status": "ok",
  "data": {
    "categories": [/* category objects */],
    "items": [/* item objects */],
    "resources": [/* resource objects */]
  }
}

Error Response

{
  "status": "error",
  "error": "Error message here"
}

Common errors:

• MEMU_API_KEY environment variable not set - User needs to configure API key
• API request failed - Network issue or API down
• Empty results - No matching memories found (not an error, just empty data arrays)

Practical Examples

Example 1: Retrieving Code Preferences (RAG)

User asks: "What languages do I prefer?"

Process:

scripts/memu.py retrieve --query "What programming languages does the user prefer?" --method rag

Response:

{
  "status": "ok",
  "data": {
    "items": [
      {"memory_type": "preference", "summary": "Prefers TypeScript over JavaScript"},
      {"memory_type": "preference", "summary": "Uses Python for scripting"}
    ]
  }
}

Your response to user: "Based on our previous discussions, you prefer TypeScript over JavaScript for application development, and you use Python for scripting tasks."

Example 2: Complex Workflow Question (LLM)

User asks: "How do I usually approach refactoring?"

Process:

scripts/memu.py retrieve --query "How does the user approach refactoring code?" --method llm

Response:

{
  "status": "ok",
  "data": {
    "categories": [
      {
        "name": "Refactoring Patterns",
        "summary": "User follows a systematic approach: write tests first, refactor in small steps, run tests after each change. Prefers extracting functions over inline complexity."
      }
    ]
  }
}

Your response to user: "You typically follow a systematic refactoring approach: you start by writing tests to ensure behavior is preserved, then refactor in small incremental steps, running tests after each change. You prefer extracting functions rather than leaving complex logic inline."

Example 3: No Results

User asks: "What do you know about my deployment process?"

Process:

scripts/memu.py retrieve --query "What is the user's deployment process?" --method rag

Response:

{
  "status": "ok",
  "data": {
    "categories": [],
    "items": [],
    "resources": []
  }
}

Your response to user: "I don't have any stored information about your deployment process yet. We haven't discussed this in previous sessions."

Quick Reference

Scenario	Method	Query Example
Specific preference	RAG	"What editor does the user prefer?"
Yes/no fact	RAG	"Does the user use TypeScript?"
Open-ended pattern	LLM	"How does the user structure React components?"
Complex workflow	LLM	"What is the user's testing strategy?"
No prior discussion	Either	(Will return empty results)

Remember: Conversations are auto-memorized at session end, so knowledge accumulates over time.