eternnoir Memory Recall
Search and retrieve relevant memories from the memory files.
When to Use
Trigger when:
- The current topic might relate to past discussions
- The user references something from a previous conversation
- You're uncertain about context that may have been discussed before
- The user asks "do you remember..." or similar queries
Do NOT guess or assume — search memory files first, then respond with confidence.
Note: memory_preferences.md is already auto-loaded into context via the @ reference in CLAUDE.md — you usually do not need to re-read it here unless the user just updated it.
Workflow
1. Read Configuration
Read .claude/memory-settings.json to get the configured file names.
2. Extract Search Keywords
From the current conversation context, identify:
- Key topics, names, technologies, or concepts
- Time references ("last week", "when we discussed X")
- Specific terms the user mentioned
3. Search in Parallel
Execute searches across all three files simultaneously:
Preferences File
- Read the full file (it's small, designed to be loaded entirely)
- Extract relevant sections
Conversations File
- Use Grep to search for keywords
- Results are in reverse chronological order (newest first)
- Look for matching
### YYYY-MM-DDentries and their content
Long-Term Memory File
- Use Grep to search for keywords
- Check across all subsections (milestones, decisions, lessons, etc.)
4. Synthesize Results
Combine findings from all files into a coherent response:
- Prioritize the most relevant and recent information
- Cross-reference between files (e.g., a preference change might relate to a conversation)
- Present information naturally — don't list "found in file X"
5. Response Style
Integrate recalled memories naturally into your response. Examples:
Good: "Based on our previous discussion, you decided to use PostgreSQL for this project. You also mentioned preferring connection pooling via PgBouncer."
Bad: "I searched memory_conversations.md and found an entry from 2025-01-15 that mentions PostgreSQL. I also found in memory_longterm.md under Architecture Decisions..."
The user should feel like you genuinely remember, not like you're reading from a database.