recall

name	recall
user-invocable	true
description	Search past Claude Code sessions and Obsidian notes to recover decisions, debugging insights, and context. Use when asked to recall, find, or search past sessions/notes, or when user says "recall". Triggers on "recall", "past session", "did I", "how did I", "find the session where".
allowed-tools	Bash(qmd *), Bash(grep *), Bash(ls *), Bash(find *), Read, Grep, Glob
effort	medium

recall — Search Past Sessions & Notes

Search indexed Claude Code and Codex CLI sessions (local and remote machines) and Obsidian research notes.

Context

• Available collections: !'qmd collection list 2>/dev/null || echo "QMD not available"'

Collections

Sessions are organized by agent and host under ~/.recall/.

Directory	Content
claude-code-local/	Claude Code sessions — this machine
claude-code-<host>/	Claude Code sessions — remote machine
codex-local/	Codex CLI sessions — this machine
codex-<host>/	Codex CLI sessions — remote machine
obsidian-<vault>/	Obsidian research notes (optional)

Each machine gets one Claude Code directory and one Codex directory. QMD collection names are user-defined when running qmd collection add — always check qmd collection list for the actual names to pass to -c. Omit -c to search all collections at once.

Search Modes

1. Topic Search (default)

QMD provides three search commands. Pick the right one for the query:

Command	Type	Best for	Example
qmd query	Hybrid (BM25 + semantic)	General queries, best overall ranking — use for ~80% of searches	"gradient accumulation", "QMD video project status"
qmd search	BM25 (keyword)	Exact strings, known identifiers, error messages	"tmux prefix", "CUDA error", "FileNotFoundError"
qmd vsearch	Semantic (embeddings)	Conceptual/fuzzy queries where exact words won't appear in the text	"times I was stuck on a bug", "ideas I never followed up on"

# BM25 — fast, deterministic, best for known keywords
qmd search "<query>" -n 5
qmd search "<query>" -c local -n 5

# Semantic — finds meaning even without exact word matches
qmd vsearch "<query>" -n 5
qmd vsearch "<query>" -c vault -n 5

# Hybrid — combines both, best ranking
qmd query "<query>" -n 5
qmd query "<query>" -c <collection> -n 5

Multi-query expansion: For vague or conceptual queries, expand into 2-3 related searches with different phrasings and merge results. Increase -n beyond 5 (up to 15) for broad recall. Example — user asks "when was I happy":

qmd vsearch "happy, grateful, excited" -c vault -n 10
qmd vsearch "energy, great day, feeling good" -c vault -n 10
qmd vsearch "satisfaction, accomplishment, shipped" -c vault -n 10

For BM25 synonym expansion (exact keywords with variation):

qmd search "CUDA error" -n 10
qmd search "GPU out of memory" -n 10
qmd search "RuntimeError device" -n 10

For hybrid broad recall:

qmd query "debugging training failures" -n 15

Merge and deduplicate before synthesizing.

2. Temporal Search — Structured Timeline

For date-based queries (recall yesterday, recall what did I do on Feb 20):

1. Resolve the date(s) — "yesterday" → YYYY-MM-DD, "last week" → date range
2. Find all sessions for that date across all machines:

   find ~/.recall/claude-code-* -name "YYYY-MM-DD-*.md" | sort
   

3. For each session, extract from frontmatter: slug, project, host, message-count, start-time, end-time, files-modified
4. Present as a structured timeline:

   ## 2026-02-20 — 36 sessions
   
   | Time | Machine | Project | Msgs | Files | Slug | Summary |
   |------|---------|---------|------|-------|------|---------|
   | 09:15–09:45 | local | dotfiles | 8 | 3 | tmux-setup | Set up tmux config |
   | 10:00–11:30 | remote | my-project | 24 | 7 | grad-accum-fix | Fixed gradient accumulation bug |
   | ... | | | | | | |
   

5. Use the slug from frontmatter as a quick label; read the first user message to generate a one-line summary
6. Show files-modified count from frontmatter (list full paths only if the user asks to drill down)
7. Include the file path of each session markdown so the user can drill into any row

3. Project Search

Search within a specific project:

qmd search "<query>" -c local --path "claude-code-local/<project>/" -n 5

To list projects per machine:

ls ~/.recall/claude-code-local/
# ls ~/.recall/claude-code-<host>/  # additional machines if configured

4. File Search

Find sessions that touched a specific file:

grep -rl "filename.py" ~/.recall/claude-code-*/ | head -10

Workflow

1. Parse the user's query to determine the search mode:
   • Date mentions ("yesterday", "last week", "Feb 20") → temporal
   • Project name → project search
   • File name → file search
   • "notes about X" or "what do I know about X" → vault collection
   • Otherwise → topic search (all collections)
2. Choose the right QMD command (see topic search table above)
3. For vague/conceptual queries, run 2-3 parallel searches with different phrasings
4. Read the top results (use the Read tool on returned file paths)
5. Synthesize — don't just list results. Extract patterns, surface forgotten decisions, and highlight actionable next steps:
   • Date and project for each relevant session
   • Machine (which host the session ran on)
   • Key decisions and their rationale — quote the exact words from the session
   • Files modified if relevant
   • Patterns across sessions (e.g., "you tried approach X three times and abandoned it each time")
   • Unfinished threads — ideas mentioned but never acted on
   • End with: "What would you like to dig into?"
6. If initial results aren't relevant, try alternate queries or combine modes

Syncing Remote Machines

bash sync_remotes.sh  # run from the skill directory

Notes

• Session frontmatter: session-id, slug, project, host, cwd, git-branch, date, start-time, end-time, message-count, files-modified
• If QMD returns no results, fall back to grep:

  grep -rli "<query>" ~/.recall/claude-code-*/ ~/.recall/codex-*/ ~/.recall/obsidian-*/ | head -10