Recall Skill

Three modes: temporal (date-based session timeline), topic (BM25 search across QMD collections), and graph (interactive visualization of session-file relationships). Every recall ends with the One Thing - a concrete, highest-leverage next action synthesized from the results.

What It Does

• Temporal queries ("yesterday", "last week", "what was I doing"): Scans native Claude Code JSONL files by date. Shows a table of sessions with time, message count, and first message. Expand any session for conversation details.
• Topic queries ("QMD video", "authentication"): BM25 search across sessions, notes, and daily logs in QMD collections.
• Graph queries ("graph yesterday", "graph last week"): Generates an interactive HTML graph showing sessions as nodes connected to files they touched. Sessions colored by day, files colored by folder. Clusters reveal related work streams, shared files show cross-session dependencies.
• One Thing synthesis: After presenting results, synthesizes the single most impactful next action based on what has momentum, what's blocked, and what's closest to done. Not generic - specific and actionable.

No custom setup needed for temporal recall - every Claude Code user has JSONL files.

Usage

/recall yesterday
/recall last week
/recall 2026-02-25
/recall QMD video
/recall authentication work

Graph mode - visualize session relationships over time:

/recall graph yesterday        # what you touched today
/recall graph last week        # week overview - find clusters
/recall graph this week        # current week so far
/recall graph last 3 days      # recent activity window

Graph options: --min-files 5 for cleaner graphs (only sessions touching 5+ files), --all-projects to scan beyond current vault.

Step 1: Locate Scripts

Find the plugin's scripts directory:

PLUGIN_DIR="$(dirname "$(dirname "$(cd "$(dirname "$0")" && pwd)")")"
SCRIPTS_DIR="$PLUGIN_DIR/scripts"

If the plugin is installed via Claude Code plugin system, scripts are at: <plugin-root>/scripts/recall-day.py, <plugin-root>/scripts/session-graph.py

Fallback: search common locations:

1. ./scripts/ (relative to plugin root)
2. ~/.claude/skills/recall/scripts/ (standalone install)

Step 2: Classify Query

Parse the user's input after /recall and classify:

• Graph - starts with "graph": "graph last week", "graph yesterday", "graph today" -> Go to Step 3C
• Temporal - mentions time: "yesterday", "today", "last week", "this week", a date, "what was I doing", "session history" -> Go to Step 3A
• Topic - mentions a subject: "QMD video", "authentication", "lab content" -> Go to Step 3B
• Both - temporal + topic: "what did I do with QMD yesterday" -> Go to Step 3A first, then scan results for the topic

Step 3A: Temporal Recall (JSONL Timeline)

Run the recall-day script:

python3 "$SCRIPTS_DIR/recall-day.py" list DATE_EXPR

Replace DATE_EXPR with the parsed date expression. Supported:

• yesterday, today
• YYYY-MM-DD
• last monday .. last sunday
• this week, last week
• N days ago, last N days

Options:

• --min-msgs N - filter noise (default: 3)
• --all-projects - scan all projects, not just current vault

Present the table to the user. If they pick a session to expand:

python3 "$SCRIPTS_DIR/recall-day.py" expand SESSION_ID

This shows the conversation flow (user messages, assistant first lines, tool calls).

Step 3B: Topic Recall (QMD BM25 with Query Expansion)

BM25 is keyword-based - it only finds exact word matches. The user's recall of a topic often uses different words than the session itself. Fix: expand the query into 3-4 keyword variants covering synonyms and related phrasings.

Step 3B.1: Expand query into variants. Generate 3-4 alternative phrasings that someone might use for the same topic. Example:

• User says "disk clean up" -> variants: "disk cleanup free space", "large files storage", "delete cache bloat GB", "free up computer space"

Step 3B.2: Run ALL variants across ALL collections in parallel (fast, ~0.3s each):

qmd search "VARIANT_1" -c sessions -n 5
qmd search "VARIANT_2" -c sessions -n 5
qmd search "VARIANT_3" -c sessions -n 5
qmd search "VARIANT_1" -c projects -n 5
qmd search "VARIANT_2" -c projects -n 5
qmd search "VARIANT_1" -c source -n 5
qmd search "VARIANT_2" -c source -n 5

Run all collection variants in parallel.

Step 3B.3: Deduplicate results by document path. If same doc appears in multiple searches, keep the highest score. Present top 5 unique results.

Step 3C: Graph Visualization

Strip "graph" prefix from query to get the date expression. Run:

python3 "$SCRIPTS_DIR/session-graph.py" DATE_EXPR

Options:

• --min-files N - only show sessions touching N+ files (default: 2, use 5+ for cleaner graphs)
• --min-msgs N - filter noise (default: 3)
• --all-projects - scan all projects
• -o PATH - custom output path (default: /tmp/session-graph.html)
• --no-open - don't auto-open browser

Opens interactive HTML in browser. Session nodes colored by day, file nodes colored by folder. Tell the user the node/edge counts and what to look for (clusters, shared files).

Step 4: Fetch Full Documents (Topic path only)

For the top 3 most relevant results across all collections, get the full document:

qmd get "qmd://collection/path/to/file.md" -l 50

Use the paths returned from Step 3B searches. The -l 50 flag limits to 50 lines.

Step 5: Present Structured Summary

For temporal queries: Present the session table and offer to expand any session.

For topic queries: Organize results by collection type:

Sessions (session logs + Claude memory)

• What was worked on related to this topic
• Key dates and decisions
• Current status or next steps

Projects (work projects)

• Relevant documentation, specs, designs

Source (personal projects)

• Personal project docs and notes

Keep this concise - it's context loading, not a full report.

Step 6: Synthesize "One Thing"

After presenting recall results, synthesize the single highest-leverage next action.

How to pick the One Thing:

1. Look at what has momentum - sessions with recent activity, things mid-flow
2. Look at what's blocked - removing a blocker unlocks downstream work
3. Look at what's closest to done - finishing > starting
4. Weigh urgency signals: deadlines in session titles, "blocked" status, time-sensitive content

Format: Bold line at the end of results:

One Thing: [specific, concrete action]

Good examples:

• One Thing: Finish the QMD video outline - sections 3-5 are drafted, just needs the closing CTA
• One Thing: Unblock the lab deploy - the DNS config is the only remaining blocker

Bad examples (too generic):

• "Continue working on the video"
• "Pick up where you left off"

If the recall results don't have enough signal, skip it and ask "What would you like to work on from here?" instead.

Fallback: No Results Found

No results found for "QUERY". Try:
- Different search terms
- Broader keywords / different date range
- --min-msgs 1 to include short sessions

Notes

• Temporal queries go through recall-day.py (native JSONL, no QMD needed)
• Graph queries go through session-graph.py (NetworkX + pyvis)
• Topic queries use BM25 (qmd search) NOT hybrid (qmd query) - 53x faster
• Run all 3 collection searches in parallel to keep response time fast