kmshihab7878 Cross-Session Recall
Search across all past conversations and memory to find context from any previous session.
Search Strategy (3-Layer)
Layer 1: claude-mem Semantic Search (Fastest, Most Relevant)
Use the mcp__plugin_claude-mem_mcp-search__search tool:
search(query="<user's search term>", limit=10)
Then fetch details for matching IDs:
get_observations([id1, id2, ...])
This searches the vector-indexed semantic memory (ChromaDB) for conceptually similar content.
Layer 2: Memory File Search (Structured Facts)
Search the memory index at ~/.claude/projects/-Users-<YOUR_USER>/memory/MEMORY.md:
- Grep MEMORY.md for keyword matches
- Read matching memory files for full context
- Check confidence-scored facts table for relevant entries
Layer 3: Conversation History Search (Raw Transcripts)
Search session history files:
# Find all history files across projects
find ~/.claude/projects/ -name "history.jsonl" -type f 2>/dev/null
# Search for keyword matches (most recent first)
grep -l "<search_term>" ~/.claude/projects/*/history.jsonl 2>/dev/null
Note: history.jsonl files can be very large. Use grep to find matching lines, don't read entire files.
Usage
/recall <search query>
Examples
/recall security audit results from last week/recall what did we decide about the database migration/recall aster trading configuration/recall coremind architecture decisions
Output Format
For each result, show:
- Source: claude-mem | memory-file | history
- Date: When it was recorded
- Relevance: How well it matches the query
- Content: The actual recalled information
- Context: Link to the memory file or session where it originated
Rules
- Always start with claude-mem semantic search (fastest, best relevance)
- Fall back to memory files if semantic search returns <3 results
- Only search history.jsonl as last resort (expensive, noisy)
- Deduplicate results across layers
- Sort by relevance, then recency
- Show max 10 results unless user asks for more
- If the user says to ignore memory, skip all layers and say so