Summary

Active Memory reliably times out and can leave the local gateway CPU-bound when configured to use cerebras/gpt-oss-120b, even though the same Cerebras key/model works through direct /v1/chat/completions calls, including streaming and tool calls.

This looks related to reasoning-model handling inside the embedded Active Memory sub-agent path, not a Cerebras auth/model-access failure.

Environment

• OpenClaw: 2026.4.26 (be8c246)
• OS: macOS / Darwin 25.4.0 arm64
• Node: v22.22.0
• Install method: npm/Homebrew global install
• Gateway mode: local LaunchAgent, port 18789
• Channel: Discord channel session
• Active Memory model: cerebras/gpt-oss-120b
• Cerebras provider API: openai-completions

Active Memory config used for reproduction

{
  "plugins": {
    "entries": {
      "active-memory": {
        "enabled": true,
        "config": {
          "enabled": true,
          "agents": ["main", "kong"],
          "allowedChatTypes": ["direct", "group", "channel"],
          "queryMode": "message",
          "promptStyle": "balanced",
          "thinking": "off",
          "maxSummaryChars": 220,
          "timeoutMs": 5000,
          "persistTranscripts": false,
          "logging": true,
          "model": "cerebras/gpt-oss-120b"
        }
      }
    }
  }
}

Steps to reproduce

1. Configure a Cerebras provider with gpt-oss-120b.
2. Configure Active Memory to use cerebras/gpt-oss-120b.
3. Enable Active Memory for an interactive persistent Discord channel session.
4. Send a normal lightweight message in the channel.
5. Watch gateway logs and process health.

Expected behavior

Active Memory should either:

• return a compact summary or NONE within the configured timeout, or
• fail fast and non-disruptively, allowing the main reply path and gateway health/admin endpoints to continue working.

Actual behavior

Active Memory starts, then times out much later than the configured timeoutMs, returns no summary, and the gateway process can become CPU-bound. During the stall, /health and websocket/admin probes may time out.

Observed logs:

2026-04-29T04:58:04.636+08:00 [plugins] active-memory: agent=main session=agent:main:discord:channel:1468250966536228866 activeProvider=cerebras activeModel=gpt-oss-120b start timeoutMs=5000 queryChars=882
2026-04-29T04:58:29.215+08:00 [plugins] active-memory: agent=main session=agent:main:discord:channel:1468250966536228866 activeProvider=cerebras activeModel=gpt-oss-120b done status=timeout elapsedMs=24580 summaryChars=0

Process/health symptoms observed after reproduction:

openclaw-gateway CPU: ~100%
curl http://127.0.0.1:18789/health: timed out after 5-8s, or returned only after several seconds
openclaw gateway status: websocket probe timeout

Disabling only plugins.entries.active-memory.config.enabled=false was not enough to return the gateway to a clean state. Disabling the plugin entry itself and restarting restored stability:

{
  "plugins": {
    "entries": {
      "active-memory": {
        "enabled": false,
        "config": {
          "enabled": false
        }
      }
    }
  }
}

After disabling the plugin entry and restarting:

Gateway health: ~2.5ms
openclaw-gateway CPU: ~0.1%

Direct Cerebras API checks

The same configured Cerebras API key/model succeeds outside the Active Memory embedded path:

• Plain /v1/chat/completions: HTTP 200, visible content returned.
• Streaming /v1/chat/completions with tools: HTTP 200, tool calls returned.
• Tool-result followup: HTTP 200, visible content returned.

Example observed direct checks:

gpt-oss-120b plain chat: 200 OK, content="ok", ~639ms
gpt-oss-120b streaming tools: 200 OK, finish_reason="tool_calls", ~432ms
gpt-oss-120b tool-result followup: 200 OK, visible content returned, ~370ms

This suggests the key has real chat/completions access and the model can handle tool-call shape at the API layer.

Notes / suspected cause

gpt-oss-120b returns reasoning side-channel content even for simple requests. My suspicion is that the Active Memory embedded sub-agent path is treating the reasoning/tool activity as hidden thinking while waiting/retrying for visible summary text, then does not abort cleanly when the Active Memory timeout fires.

This looks related to, but not identical to:

• Active Memory timeout with MiniMax-M2.7: model fails to respond within 15s timeout #66804 Active Memory timeout with a reasoning model
• [Bug]: OpenClaw 2026.3.12: Thinking models (reasoning: true) timeout across multiple providers (zai/glm-5, minimax-portal) #45681 reasoning models timing out across providers
• [Bug]: active-memory sub-agent receives empty tools array, memory_search/memory_get never called #68004 Active Memory tools handling regression
• [Bug]: GPT-OSS-120B function calling fails with JSON parse error in OpenClaw #9956 GPT-OSS tool/function parsing behavior

The main difference here is that the docs recommend cerebras/gpt-oss-120b for Active Memory, and direct Cerebras chat/tool calls work, but the embedded Active Memory path times out and can destabilize the gateway.

Workaround

Disable Active Memory entirely, or avoid cerebras/gpt-oss-120b as the Active Memory model until the embedded Active Memory path handles reasoning-model responses/timeouts cleanly.