/recall — Search Past Claude & Codex Sessions

Search all past Claude Code and Codex sessions using full-text search with BM25 ranking.

Usage

python3 ~/.claude/skills/recall/scripts/recall.py QUERY [--project PATH] [--days N] [--source claude|codex] [--limit N] [--reindex]

Examples

# Simple keyword search
python3 ~/.claude/skills/recall/scripts/recall.py "bufferStore"

# Phrase search (exact match)
python3 ~/.claude/skills/recall/scripts/recall.py '"ACP protocol"'

# Boolean query
python3 ~/.claude/skills/recall/scripts/recall.py "rust AND async"

# Prefix search
python3 ~/.claude/skills/recall/scripts/recall.py "buffer*"

# Filter by project and recency
python3 ~/.claude/skills/recall/scripts/recall.py "state machine" --project ~/my-project --days 7

# Search only Claude Code sessions
python3 ~/.claude/skills/recall/scripts/recall.py "buffer" --source claude

# Search only Codex sessions
python3 ~/.claude/skills/recall/scripts/recall.py "buffer" --source codex

# Force reindex
python3 ~/.claude/skills/recall/scripts/recall.py --reindex "test"

Query Syntax (FTS5)

• Words: bufferStore — matches stemmed variants (e.g., "discussing" matches "discuss")
• Phrases: "ACP protocol" — exact phrase match
• Boolean: rust AND async, tauri OR electron, NOT deprecated
• Prefix: buffer* — matches bufferStore, bufferMap, etc.
• Combined: "state machine" AND test
• CJK: issue化 — automatically uses trigram matching for Japanese/Chinese/Korean text

After Finding a Match

To resume a session, cd into the project directory and use the appropriate command:

# Claude Code sessions [claude]
cd /path/to/project
claude --resume SESSION_ID

# Codex sessions [codex]
cd /path/to/project
codex resume SESSION_ID

Each result includes a File: path. Use it to read the raw transcript (auto-detects format):

python3 ~/.claude/skills/recall/scripts/read_session.py <File-path-from-result>

If results are missing File: paths, run --reindex to backfill.

Notes

• Index is stored at ~/.recall.db (SQLite FTS5, auto-migrated from ~/.claude/recall.db)
• Indexes both ~/.claude/projects/ (Claude Code) and ~/.codex/sessions/ (Codex)
• First run indexes all sessions (a few seconds); subsequent runs are incremental
• Only user and assistant messages are indexed (tool calls, thinking blocks, state snapshots skipped)
• Results show [claude] or [codex] tags to indicate the source
• Dual-table FTS: English queries use Porter stemming, CJK queries use trigram matching
• Upgrading from 0.2.x: run --reindex once to build the CJK index