Claude Recall

Turn your repetitive Claude Code workflows into reusable skills — automatically.

中文文档 • Quick Start • FAQ

---

You use Claude Code every day. You deploy the same way, debug the same way, restart dev servers the same way. Claude Recall watches silently, finds those repeated workflows, and turns them into reusable skills you can invoke forever.

How It Works

 You use Claude Code normally
          │
          ▼
 [Hook] log-operations.py silently records every tool call
          │
          ▼
 After a few sessions, you run /analyze-patterns
          │
          ▼
 Python pre-analyzer crunches logs → Claude reads the summary (not raw logs)
          │
          ▼
 You pick which patterns to save as /skills
          │
          ▼
 New skills are written to ~/.claude/ — ready to use

Three components, zero friction:

Component	What it does
log-operations.py	PostToolUse hook — silently appends a JSON summary after every tool call. Auto-rotates at 10 MB.
pre-analyze.py	Python script — crunches raw logs into structured summaries. ~80% token savings vs reading raw JSONL.
analyze-patterns	Skill — reads the pre-analyzed summary, detects patterns across 3+ sessions, generates reusable skill files.

Quick Start

git clone https://github.com/zchdu/claude-recall.git
cd claude-recall
./install.sh

That's it. Use Claude Code normally for a few sessions, then run:

/analyze-patterns

What Gets Logged

Each tool call produces one compact JSONL entry:

{
  "ver": 2,
  "ts": "2026-03-02T09:50:10Z",
  "sid": "e0af7856-2df8-48",
  "tool": "Bash",
  "input": {"command": "npm test", "description": "Run tests"},
  "cwd": "/home/user/my-project",
  "tuid": "toolu_01ABcDeF12",
  "res": {"success": true, "exit_code": 0}
}

Feature	Detail
Format	v2 — includes tool response summary (res) and tool use ID (tuid)
Truncation	Strings over 300 chars trimmed, responses summarized to 500 chars
Rotation	Auto-rotates at 10 MB, keeping the newer half
Location	~/.claude/tool_logs/operations.jsonl

What /analyze-patterns Does

Step	Description
1. Pre-analyze	Python script crunches logs into statistics + patterns (~10 KB summary)
2. Statistics	Total records, session count, tool frequency top 5, most common commands and directories
3. Pattern detection	Finds repeated commands, sequences (2–3 steps), and workflow patterns across 3+ sessions
4. Suggestions	For each pattern: name, frequency, steps, parameterizable parts, and a ready-to-use .md skill
5. Creation	You choose which to save → written to ~/.claude/commands/ → immediately available

Example Output

=== Pre-analysis Summary ===
Records: 853 | Sessions: 12 | Span: 7d
Tool Frequency: Bash (321, 38%), Read (253, 30%), Edit (124, 15%)
Existing Skills: restart-dashboard, dashboard-logs

Detected Patterns:
  Repeated: `ssh kubao "cd /opt/X && ..."` — 5 sessions
  Sequence: Edit(config) → Bash(cargo build) → Bash(deploy) — 4 sessions

=== Pattern: Deploy to Production ===
- Overlap: None
- Recommendation: Create new skill
- Frequency: 4 sessions
- Steps: rsync → ssh restart → tail logs

Which patterns to save? > 1

✓ Created ~/.claude/commands/deploy-prod.md
  Invoke with: /deploy-prod

Manual Installation

Click to expand manual steps

1. Copy files

mkdir -p ~/.claude/{hooks,scripts,skills/analyze-patterns}
cp hooks/log-operations.py ~/.claude/hooks/
cp scripts/pre-analyze.py ~/.claude/scripts/
cp skills/analyze-patterns/SKILL.md ~/.claude/skills/analyze-patterns/
chmod +x ~/.claude/hooks/log-operations.py ~/.claude/scripts/pre-analyze.py

2. Register the hook

Add this to your ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "python3 $HOME/.claude/hooks/log-operations.py",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

If you already have a hooks section, merge the PostToolUse entry into it.

File Structure

~/.claude/
├── hooks/
│   └── log-operations.py       # PostToolUse hook (v2 format)
├── scripts/
│   └── pre-analyze.py          # Log pre-analyzer
├── skills/
│   └── analyze-patterns/
│       └── SKILL.md            # /analyze-patterns skill
├── commands/
│   └── analyze-patterns.md     # Legacy skill (backward compat)
├── tool_logs/
│   └── operations.jsonl        # Auto-generated log
└── settings.json               # Hook registration

FAQ

Does it slow down Claude Code?

No. The hook runs with a 5-second timeout and typically finishes in under 10 ms. It is a single JSON append.

How much disk space does it use?

The log auto-rotates at 10 MB. Typical usage produces roughly 1 MB per week.

Does it capture sensitive data?

File contents and diffs are truncated to 300 characters. Tool responses are summarized to 500 characters (key fields only). You can review ~/.claude/tool_logs/operations.jsonl at any time.

Can I log only specific tools?

Yes. Change the matcher field in settings.json. For example, "matcher": "Bash|Edit" logs only Bash and Edit calls.

What if I already have a settings.json?

install.sh detects existing settings and safely merges the hook config without overwriting your file.

How do I uninstall?

Run ./uninstall.sh. It removes all components while preserving your other settings.

Requirements

• Claude Code with hooks support
• Python 3.8+

Contributing

Contributions are welcome. Please see CONTRIBUTING.md for guidelines.

License

MIT