Problem

A single runaway session can consume unbounded memory and API costs with no automatic intervention. The gateway process will crash (or OOM-kill) before any safety mechanism triggers.

Impact (real incident, Feb 14 2026)

One session grew from 800MB to 3,021MB RSS in 20 minutes:

The server has 6GB heap allocated. One session consumed 50% of total heap.

Estimated API cost for this single session: $150+ in ~2 hours (1,535 polls × 187k cached tokens each).

Proposed Solution

Add configurable per-session resource limits:

1. Memory guard

{
  "agents": {
    "defaults": {
      "session": {
        "maxRssMb": 2048,
        "maxRssAction": "compact-or-kill"
      }
    }
  }
}

When a session pushes total gateway RSS above threshold:

• First: force compaction on the largest session
• If RSS still high after compaction: archive the session and start fresh
• Notify the user on their configured channel

2. Session cost tracking

Track per-session: input tokens, output tokens, cache reads. Surface in session_status. Alert if a single session exceeds a configurable threshold (e.g., $50).

3. Session message count guard

If a session exceeds N messages (e.g., 500) without compaction, force compaction. The Feb 14 session hit 1,535 messages — this should never happen.

Why This Matters

On a resource-constrained server (common for self-hosted OpenClaw), one bad session can take down the entire gateway, affecting all agents and all channels. There is no isolation between sessions for memory.

Environment

• OpenClaw 2026.2.13
• Ubuntu 24.04, 6GB Node.js heap (--max-old-space-size=6144)
• Gateway running as single Node.js process serving all agents/channels