Deep Session Search

This skill helps recover decision context (domains, PRs, deploys, commands, and error trails) from historical OpenClaw sessions when those details were discussed but never promoted into long-term memory files.

Why this exists

Sometimes important details are discussed in earlier sessions but never distilled into MEMORY.md or daily notes.

This skill gives the agent a reliable fallback path:

1. Try normal memory recall first.
2. If recall is missing/unclear, search session logs under ~/.openclaw/agents/main/sessions.
3. Extract high-signal conversation/tool evidence without dumping noisy artifacts.

What it does

• Deep session grep over *.jsonl* logs
• Noise control to avoid dependency/build blobs
• Targeted narrowing by exact domain, PR number, repo, deployment id, etc.
• Evidence-first summaries with path + relevant context

Typical trigger

Use when the user asks things like:

• “What did we do with Cloudflare today?”
• “Check earlier sessions for that domain/config”
• “Find the command we used before”

…especially when memory search has low confidence.

Search pattern

Prefer ag for fast uncompressed session log scans:

ag -n --nobreak --depth 1 -G '*.jsonl' "<search-term-1>|<search-term-2>" \
  ~/.openclaw/agents/main/sessions

For compressed/rotated logs, use one of these:

# ripgrep with compressed-file support
rg -n --search-zip --max-depth 1 "<search-term-1>|<search-term-2>" \
  ~/.openclaw/agents/main/sessions -g '*.jsonl*'

# or gzip-specific fallback
find ~/.openclaw/agents/main/sessions -maxdepth 1 -name '*.jsonl.gz' \
  -exec zgrep -HnE "<search-term-1>|<search-term-2>" {} +

Note: these examples use regex patterns. If your search terms are literals (like domains), either escape regex special characters or use fixed-string modes (for example: ag -Q, rg -F, zgrep -F).

Then narrow to exact needles for stronger evidence.

Outcome format

The skill returns concise findings with:

• whether hits were found
• strongest 1–3 examples
• where they were found (path/context)
• confidence note if results are partial/noisy