Main thread CPU pinning at 100%: aggressive SQLite polling in memory-tdai pipeline

[fdefilippo]

Description

When running as the context engine in OpenClaw v2026.6.1, the memory-tencentdb plugin (v0.3.6) causes the main Node.js thread to be pinned at 99.9% CPU usage permanently. Total CPU consumption across worker threads reaches ~150% on a single core.

Environment

• OS: Oracle Linux Server 10.1
• Node.js: v24.14.1
• OpenClaw: 2026.6.1
• Plugin: TencentDB-Agent-Memory v0.3.6
• Memory: ~1.1 GB RSS consumed by the process

Symptoms

1. Main thread is permanently at 99.9% CPU (state R - running)
2. System load average stays above 2.0 even during idle periods
3. Memory consumption stays at ~1.1GB RSS even when no user interaction is happening

Log Analysis

The memory-tdai pipeline performs frequent SQLite L1-queries across multiple commitment sessions. Each cycle shows:

1. [memory-tdai][sqlite] [L1-query] — runs for each commitment session key every ~10-15 minutes
2. [memory-tdai] [pipeline] [L2 cold-start skip] — consistently runs and skips (wasted I/O)
3. [context-offload] [local-llm] — makes repeated calls to the configured model
4. [memory-tdai] [recall] Strategy 'hybrid' requested but EmbeddingService not available, falling back to keyword — embedding service is not configured

Key observations

• SQLite L1-queries run even when there are no new messages from any user
• Multiple commitment sessions are being polled constantly (10+ different UUIDs observed)
• The L2 pipeline cold-start runs every cycle even though it always skips because minInterval hasn't passed
• No embedding service available causes fallback to keyword search on every recall

Expected behavior

The plugin should not consume CPU during idle periods. SQLite polling should have a configurable interval (or be event-driven rather than poll-driven). The L1 and L2 pipeline should not run when there are no new interactions to process.

Suggested investigation areas

1. Polling interval: L1-query appears to run aggressively for each commitment session even when idle. Consider adding a configurable polling interval or switching to a trigger-based approach.
2. Event loop blocking: Verify that all SQLite operations use async/promises and do not block the Node.js event loop.
3. Cold-start cycles: L2 pipeline cold-start runs every cycle and always skips. Consider checking minInterval before triggering the pipeline query.
4. Embedding service: When embedding service is unavailable, cache the strategy decision rather than falling back on every recall.

[Maxwell-Code07]

Thanks for the bug report! We've received your report about 100% CPU pinning and will investigate internally. We'll get back to you once we have results. Thank you for the detailed environment information!