Compaction Context Recovery Plugin for OpenClaw

Preserves recent conversation context across compaction cycles, solving the "I don't know what 'all three' refers to" problem.

Installation

openclaw plugins install compaction-context

Then restart your gateway.

The Problem

When OpenClaw compacts a long session to free up context space, the agent loses access to recent conversation details. You might ask a follow-up question and get:

"Hey — context got wiped in a compaction and the summary didn't survive. I don't have what 'all three' refers to."

The Solution

This plugin hooks into the compaction lifecycle:

1. Before compaction (before_compaction hook):

   • Reads the current session's JSONL transcript
   • Extracts the last N user/assistant messages
   • Writes them to RECENT.md in the workspace
   • Sets a .compaction-recovery-pending flag

2. After compaction, on next turn (before_agent_start hook):

   • Checks if the flag exists (meaning compaction just happened)
   • If yes, reads RECENT.md and injects it as prependContext
   • Clears the flag so subsequent turns don't re-inject

Configuration

{
  "plugins": {
    "entries": {
      "compaction-context": {
        "enabled": true,
        "config": {
          "messageCount": 20,
          "maxCharsPerMessage": 500
        }
      }
    }
  }
}

| Option | Default | Description | |--------|---------|-------------| | messageCount | 20 | Number of recent messages to preserve | | maxCharsPerMessage | 500 | Truncate long messages to this limit |

Files Created

• RECENT.md — The preserved context (overwritten each compaction)
• .compaction-recovery-pending — Flag file (deleted after injection)

Token Budget

With defaults (20 messages × 500 chars), worst case is ~10K characters (~2.5K tokens) injected once after compaction. In practice it's usually much less since most messages are short.

Why This Approach?

• No constant overhead — Only runs when compaction actually happens
• Uses existing hooks — No custom OpenClaw patches needed
• Automatic — No manual intervention required
• Configurable — Tune message count and truncation to your needs

License

MIT