Description

After upgrading from 2026.2.26 to 2026.3.1, the pre-compaction memory flush fires on every single user message in TUI sessions, even seconds apart. Previously (2026.2.26), compaction was spread hours apart — now it triggers on nearly every turn.

Steps to Reproduce

1. Run openclaw tui on v2026.3.1
2. Have a session with moderate context (heartbeats, tool calls, etc.)
3. Once context grows, the Pre-compaction memory flush system message fires after every user message

Expected Behavior

Compaction should trigger occasionally when context is genuinely full, not on every turn. On v2026.2.26, the same session type saw ~3 compactions per day. On v2026.3.1, we hit 6+ compactions in a single session within ~1 hour.

Actual Behavior

• Memory flush system message appears after every single user message
• session_status shows 6 compactions with context at only 24% (49k/200k)
• Logs confirm: yesterday (v2026.2.26) had 3 compaction starts all day; today (v2026.3.1) fires continuously

Additional Context

• Related: Pre-compaction memoryFlush: larger default threshold, workspace-level flush instructions, and post-compaction audit #31435 (larger default threshold), TUI: Pre-compaction memory flush system messages visible to user #32057 (TUI showing flush messages), [Bug]: TUI displays system-generated messages (heartbeats, memory flushes, NO_REPLY) in chat history #23979 (TUI system message visibility)
• Config: contextPruning: { mode: "cache-ttl", ttl: "1h" }, compaction: { mode: "safeguard" }
• Model: claude-opus-4-6 (200k context)
• The flush itself works correctly (agent saves to memory files) — the issue is the trigger threshold is far too aggressive post-upgrade

Environment

• OpenClaw: 2026.3.1
• Previous version: 2026.2.26 (no issue)
• OS: macOS (arm64)
• Node: v25.5.0
• Channel: TUI (webchat)