jimweller STARTER_CHARACTER = 🗃️
Transcript Search
Search the global session memory database for past conversations using hybrid FTS5 + vector retrieval.
When to use
- User asks about something from a previous session or earlier in a long session
- User references a past decision, error, action, or discussion no longer in context
- User asks "what did we do about X", "when did we fix Y", "what error did we get"
Database location
~/.claude/session_memory.db (global, all projects)
Retrieval tool
TR=~/.config/dotfiles/submodules/total-recall
$TR/.venv/bin/python $TR/retrieve.py "QUERY" --db ~/.claude/session_memory.db
Default output is --format context which produces XML-tagged conversation
excerpts ready to read directly.
Common flags
| Flag | Purpose |
|---|---|
--project NAME |
Restrict to a specific project |
--top-k N |
Number of seed matches before expansion (default: 5) |
--window N |
Context window +/- N around each match (default: 2) |
--budget N |
Max chunks in final output (default: 20) |
--tokens N |
Max tokens in final output (default: 6000) |
--session ID |
Restrict to a specific session_id |
--no-tools |
Exclude tool_use/tool_result chunks |
--no-vectors |
FTS5 keyword search only (skip embeddings) |
--format json |
JSON array output for programmatic use |
--roles user,assistant |
Filter by role |
Search procedure
- Run retrieve.py with the user's query
- Read the returned context block
- If results are too broad, narrow with
--projector--session - If results are too sparse, increase
--top-kor--budget - For keyword-exact searches, use
--no-vectorsto restrict to FTS5
Tips
- The retrieval pipeline uses RRF fusion of keyword + vector search, window expansion, ancestor backtracking, and cross-session semantic links
- Use
--no-toolswhen searching for decisions or discussion (skips noisy tool output) - Use
--format jsonwhen you need to process results programmatically