Problem

When users send images via Discord/Telegram, OpenClaw:

1. Saves the image file to ~/.openclaw/media/inbound/ ✓
2. Also embeds the full base64 data in the session .jsonl file ← Problem

This causes:

• A 972KB PNG becomes 1.15MB in session history (base64 overhead)
• Session files grow rapidly (3.3MB+ with just a few images)
• Context window fills up quickly since the base64 is loaded into LLM context
• Images cannot be pruned by contextPruning since they're part of user messages

Current Behavior

// In session .jsonl
{
  "type": "message",
  "message": {
    "role": "user",
    "content": [
      {"type": "text", "text": "..."},
      {"type": "image", "data": "iVBORw0KGgo...(1MB base64)...", "mimeType": "image/png"}
    ]
  }
}

Proposed Solution

Store images as path references:

{
  "type": "image",
  "path": "/Users/.../.openclaw/media/inbound/xxx.png",
  "mimeType": "image/png"
}

When loading context:

• If image is recent (within TTL): read from disk and include
• If image is old: replace with placeholder like [Image: screenshot.png - pruned]

Additional Enhancement

After image analysis, optionally replace base64 with text summary:

{"type": "image_summary", "text": "[Screenshot showing terminal with sessions.json file size of 5.6MB]"}

Environment

• OpenClaw 2026.2.12
• macOS
• Discord channel

Impact

This would significantly reduce:

• Session file sizes
• Context consumption
• API costs (fewer tokens for images that don't need re-analysis)