context-bootstrap

Session recovery for OpenClaw agents. After /new, /reset, or heavy context compaction, your agent reads recent conversation transcripts and catches up naturally — no session injection, no database manipulation, no instability.

The Problem

When your OpenClaw agent's context window fills up and you need to reset, it loses everything — projects you were working on, decisions you made, the relationship you've been building. Starting from zero every time is exhausting.

The Solution

This skill takes a simple approach:

1. A nightly script extracts your conversations into daily transcript files
2. After a reset, your agent reads the last 2-3 days of transcripts
3. It synthesizes what happened and reports back naturally
4. You pick up where you left off

That's it. No clever hacks, no session store manipulation. Just reading and conversation.

Why not inject old messages back into the session? We tried that. Multiple times. It destabilizes the gateway, corrupts session state, creates orphaned message branches, and causes compaction failures. The "just read the transcripts" approach is boring but it works every time.

Install

Option 1: Install as a skill

# Download the .skill file from releases, then:
openclaw skill install context-bootstrap.skill

Option 2: Manual setup

Copy the files into your workspace:

your-workspace/
├── skills/
│   └── context-bootstrap/
│       └── SKILL.md
├── scripts/
│   └── extract_transcripts.sh
└── memory/
    └── transcripts/     ← created automatically

Setup

1. Configure the extraction script

Edit scripts/extract_transcripts.sh and set your agent paths:

SESSIONS_DIR="$HOME/.openclaw/agents/YOUR_AGENT_NAME/sessions"
TRANSCRIPTS_DIR="$HOME/.openclaw/workspace-YOUR_AGENT_NAME/memory/transcripts"

Or set the OPENCLAW_AGENT environment variable and it auto-detects:

export OPENCLAW_AGENT=sage
bash scripts/extract_transcripts.sh

2. Run it once to seed transcripts

chmod +x scripts/extract_transcripts.sh
bash scripts/extract_transcripts.sh

You should see transcript files appear in memory/transcripts/.

3. Set up nightly extraction

Add an OpenClaw cron job to run the extraction automatically:

Schedule: cron "0 2 * * *" (2 AM daily, adjust timezone)
Payload: agentTurn — "Run: bash scripts/extract_transcripts.sh"
Session: isolated

Or use system crontab:

crontab -e
# Add:
0 2 * * * cd /path/to/workspace && bash scripts/extract_transcripts.sh

4. Add "bootstrap" to your workflow

After any /new or /reset, just tell your agent:

bootstrap

It reads the transcripts, reads your workspace state files (HEARTBEAT.md, MEMORY.md, etc.), and tells you what it knows.

Requirements

• OpenClaw
• jq (install with apt install jq or brew install jq)
• Bash

How It Works

Session JSONL files → extract_transcripts.sh → daily .txt files → agent reads on bootstrap

OpenClaw stores conversations as JSONL files in ~/.openclaw/agents/<name>/sessions/. The extraction script parses these into clean, date-stamped text files. When the agent bootstraps, it reads the recent transcripts using the Read tool and synthesizes a summary.

The transcripts survive across session resets because they're standalone files in the workspace, not tied to any particular session.

Tips

• Transcripts too large? The agent will read only the last 1-2 days if files are big. You can also limit line count.
• Multiple agents? Each agent needs its own extraction script pointing to its own sessions directory.
• State files help: If your workspace has a HEARTBEAT.md or MEMORY.md with current priorities, the bootstrap skill reads those too for extra grounding.
• Don't over-bootstrap. Use it after resets, not on every conversation. Your agent has normal memory within a session.

License

MIT — use it however you want.

Credits

Built by Will Durocher and Kara Codex. Born from 8 hours of debugging session injection approaches before realizing the simplest solution wins every time.