| name | cw:recall |
| description | Search across all past Cursor agent sessions. Use when you need to recall previous work, find solutions to problems you've solved before, or retrieve context from past conversations. Powered by DuckDB keyword search and semantic vector search over session history. |
Recall — Cross-Session Memory
Search across all past Cursor agent sessions stored in the local DuckDB warehouse.
When to use
- "Have I worked on this before?"
- "How did I solve X last time?"
- "What did I do in project Y?"
- "Find all sessions where we discussed Z"
Finding the scripts
CURSOR_PLUGIN_ROOT should be set when invoked through the plugin system, but may be unset during development. Resolve once per session:
Use find -L so symlinked dev installs (e.g. local/cursor-warehouse → your clone) are traversed.
PLUGIN_SCRIPTS="${CURSOR_PLUGIN_ROOT:+$CURSOR_PLUGIN_ROOT/scripts}"
if [ -z "$PLUGIN_SCRIPTS" ] || [ ! -d "$PLUGIN_SCRIPTS" ]; then
PLUGIN_SCRIPTS="$(dirname "$(find -L ~/.cursor/plugins -name query.py -path '*/cursor-warehouse/*/query.py' 2>/dev/null | head -1)")"
fi
How to search
Always run BOTH searches — they complement each other.
1. Keyword search (exact substring match)
uv run --script "$PLUGIN_SCRIPTS/query.py" search "$ARGUMENTS"
2. Semantic search (meaning-based, finds related concepts)
uv run --script "$PLUGIN_SCRIPTS/vsearch.py" "$ARGUMENTS"
Finds results even when exact words don't match. Supports filters: --project X, --days N, --type message|session|message_user_query, --limit N. Omitting --type searches full message text and sessions, not message_user_query (avoids duplicate hits per turn).
Interpreting results
Keyword returns exact substring matches — high precision, low recall.
Semantic returns conceptually similar content — lower precision, high recall. Score > 0.7 = strong match, 0.5-0.7 = related.
Combine both to get a complete picture. Use session IDs to dig deeper with cw:query.
Tips
- Search for error messages, library names, patterns, concepts
- Use short, specific terms for keyword search
- Use natural language descriptions for semantic search
- Combine with
cw:query for complex lookups (e.g., "all sessions in project X that used tool Y")
