Managing Memory

Managing Memory

First-Run Setup Check

IMPORTANT: Before using memory tools, ensure the MCP server is built and configured:

# Check if built
ls ~/.claude/plugins/cache/brain-jar/shared-memory/*/dist/index.js 2>/dev/null || echo "NOT_BUILT"

If NOT_BUILT, run the setup:

node ~/.claude/plugins/cache/brain-jar/shared-memory/*/run.js &
sleep 15

Then check for Mem0 config:

cat ~/.config/brain-jar/config.json 2>/dev/null || echo "NOT_CONFIGURED"

If NOT_CONFIGURED, ask user for their Mem0 API key (get one at https://app.mem0.ai), then create config:

mkdir -p ~/.config/brain-jar
cat > ~/.config/brain-jar/config.json << 'EOF'
{
  "mem0_api_key": "USER_API_KEY_HERE",
  "default_scope": "global",
  "auto_summarize": true
}
EOF

Note: Local storage works without Mem0 config - cloud sync is optional.

After setup, user must restart Claude Code for MCP to register.

When to Store Memories

Store memories when you observe:

• Decisions - User chooses one approach over another
• Preferences - User expresses likes/dislikes about tools, patterns, or approaches
• Reactions - Strong positive or negative responses (enthusiasm, frustration)
• Context - Important background about projects, goals, or constraints

How to Write Enriched Memories

Bad (too dry):

User chose Neon for database.

Good (captures context and sentiment):

User chose Neon over Supabase for Postgres hosting - appreciated the more generous
free tier limits. Showed strong preference for managed solutions: "I'm not running
my own infra" - values simplicity and time savings over control.

Memory Format

Include:

1. The fact - What was decided/learned
2. The why - Reasoning behind it
3. The sentiment - How they felt about it (quote if memorable)
4. The implication - What this suggests about future preferences

Scope Selection

• global - Personal preferences, general learnings, cross-project patterns
• project:<name> - Specific to current project (detect from working directory)

Use global for preferences that apply everywhere. Use project: for architectural decisions, tech choices, and patterns specific to one codebase.

When to Recall Memories

Before:

• Starting a new feature (search for relevant past decisions)
• Making technology choices (search for preferences)
• Suggesting approaches (search for patterns they liked)

Use natural recall language:

• "Remember when you were working on X, you decided..."
• "You've mentioned before that you prefer..."
• "Based on your experience with Y..."

Subagent Pattern for Large Retrievals

When searching or listing memories that could return many results (10+), dispatch a Haiku subagent to process and summarize:

When to use subagent:

• Broad searches: "find all memories about preferences"
• Timeline queries: "memories from last month"
• Pattern analysis: "what decisions have I made about databases?"

How to dispatch:

Use Task tool:

• subagent_type: "general-purpose"
• model: "haiku"
• prompt: "Search user memories and provide a summary.

TASK: [what user is looking for]

STEPS:

1. Call search_memory with query '[search terms]'
2. If many results, group by theme or tag
3. Return a 3-5 sentence summary highlighting:
   • Key patterns/preferences found
   • Notable decisions or quotes
   • Relevant context for the current task

OUTPUT: Concise summary, not raw memory dumps."

Tell the user: The subagent's summary. Don't dump all raw memories into context.

When to use direct tools:

• Specific lookups: "what was the exact decision about X?"
• Small result sets: known to be 1-3 memories
• When user explicitly asks for full details

Example flow:

User: What do I usually prefer for state management?

You: Let me search your memories for state management preferences...
[Dispatch Haiku subagent]

Subagent returns:
"Found 7 memories about state management. Key patterns:
- Strong preference for Zustand over Redux (called Redux 'too much boilerplate')
- Uses React Query for server state ('keeps server/client state separate')
- Avoids global state when possible - prefers component-level state
Notable quote: 'I want the simplest thing that works'"

You: Based on your past preferences, you tend to favor Zustand for client state
and React Query for server state. You've mentioned wanting "the simplest thing
that works" - should I keep that philosophy for this feature?

Tags to Use

• preference - Likes/dislikes
• decision - Specific choices made
• architecture - System design patterns
• personality - Working style, communication preferences
• project - Project-specific context
• session-summary - End-of-session consolidation
• profile-context - Background context for profile preferences
• profile-learning - Observations that inform the user profile

Related Skills

Learning About You

For structured user profile management (name, role, tech preferences, working style), use the learning-about-you skill instead of storing as freeform memories.

Use memories for:

• Rich context and reasoning behind preferences
• Specific quotes and reactions
• Project-specific decisions
• Session summaries

Use profile for:

• Structured data (name, timezone, role)
• Tech stack preferences (languages, frameworks)
• Working style settings (verbosity, pace)
• Personal goals and interests

The profile is queryable and shared across all brain-jar plugins. Memories provide the context and "why" behind profile entries.