Description

Active Memory plugin's embedded runs (Light/REM dreaming phases) consistently timeout when using Claude Opus 4.6 via Amazon Bedrock. The gateway logs show repeated surface_error failover decisions with reason=timeout.

Environment

• OpenClaw: 2026.4.21 (f788c88)
• Node: v22.22.2
• OS: Ubuntu 24.04 (aarch64, AWS Graviton 3)
• Provider: Amazon Bedrock (us-west-2, cross-region inference)
• Model: global.anthropic.claude-opus-4-6-v1
• Plugin: active-memory (bundled)

Reproduction

1. Enable Active Memory plugin with default config
2. Use Claude Opus 4.6 as the default model
3. Wait for dreaming phases to trigger (Light/REM cycles)
4. Observe gateway logs

Observed Behavior

Embedded runs for active-memory consistently fail with timeout:

[agent/embedded] embedded run failover decision: runId=active-memory-modqufpt-9760cd06 stage=assistant decision=surface_error reason=timeout from=amazon-bedrock/global.anthropic.claude-opus-4-6-v1 profile=-

This occurs multiple times per day (10+ times in a 48-hour period). The dreaming phases never complete successfully.

One instance also showed a credential loading error:

[agent/embedded] embedded run agent end: runId=f94b865b-... isError=true error=Could not load credentials from any providers

Expected Behavior

• Embedded runs should either use a lighter/faster model for dreaming phases, or
• The timeout for embedded runs should be configurable and set appropriately for slower models, or
• Active Memory should gracefully fall back to a faster model when the primary model times out

Suggested Improvements

1. Allow configuring a separate model for Active Memory embedded runs (e.g., Sonnet for dreaming, Opus for main conversations)
2. Increase or make configurable the timeout for embedded dreaming runs
3. Add retry with model fallback before giving up

Workaround

None currently. The dreaming phases simply fail silently. Deep dreaming phase has never completed, resulting in MEMORY.md never being generated.