Summary

After upgrading from 2026.3.24 → 2026.4.15, enabling the Active Memory plugin causes every blocking sub-agent call to time out, accompanied by correlated [memory] qmd update failed (Error: qmd update timed out after 120000ms) errors in the gateway log.

Environment

• OpenClaw: 2026.4.15 (image + npm-global binary via npm install -g --prefix /data/.npm-global openclaw@latest)
• qmd: 1.0.7 (@tobilu/qmd)
• Runtime: Docker container, Node 24, Linux x86_64, CPU-only (CUDA compile fails, falls back to CPU)
• Agent: signal-chat, Signal DM direct chat type
• Active Memory config: queryMode: message, promptStyle: balanced, thinking: off, timeoutMs: 12000

Observed behavior

Every active-memory blocking sub-agent call:

1. Takes 8–12 seconds regardless of model provider
2. Times out (or returns summaryChars=0)
3. Correlates 1:1 with [memory] qmd update failed timeouts at the 120000ms cap
4. Surfaces ⚠️ Gateway: exec failed in the user-facing chat reply

Tested with both anthropic/claude-haiku-4-5 AND google/gemini-2.5-flash-lite — identical latency, suggests the bottleneck is NOT the LLM provider.

Log excerpts

[plugins] active-memory: agent=signal-chat session=agent:signal-chat:signal:direct:+REDACTED activeProvider=anthropic activeModel=claude-haiku-4-5 start timeoutMs=3000 queryChars=404
[agent/embedded] embedded run failover decision: runId=active-memory-... stage=assistant decision=surface_error reason=timeout from=anthropic/claude-haiku-4-5
[plugins] active-memory: done status=timeout elapsedMs=9839 summaryChars=0

[memory] qmd update failed (Error: qmd update timed out after 120000ms)
[memory] qmd update failed (Error: qmd update timed out after 120000ms)
(repeats per turn; 12 within a few minutes)

qmd CLI is healthy

The qmd CLI itself responds fast when invoked directly:

• qmd --help — instant
• qmd status — instant
• qmd collection add . --mask '*.md' — 31 files indexed in <1s
• qmd embed — 84 chunks with embeddinggemma-300M in 65s on CPU

So the hang is specific to the runtime path Active Memory uses to invoke qmd, not the qmd binary.

Reproduction

1. Upgrade from 2026.3.24 → 2026.4.15 (both image pull and npm install -g --prefix /data/.npm-global openclaw@latest)
2. Workspace with a non-trivial MEMORY.md (~160 lines) and large skills registry (~18 skills)
3. Enable active-memory in plugins.entries with agents: ["signal-chat"]
4. Send a DM to the agent over Signal
5. Observe ⚠️ Gateway: exec failed in the reply and qmd 120s timeouts in the gateway log

Workaround

Disabling the Active Memory plugin stops all qmd timeouts and eliminates the user-facing error surface. Memory still works via session bootstrap and post-compaction re-loads — only per-turn recall is lost.

Possible causes (speculative)

• qmd update path holds a lock or semaphore that doesn't release; subsequent calls pile up
• Embedded-run setup loads oversized context (skills registry, tool definitions) before reaching the model
• Regression from the 4.12 introduction of the blocking sub-agent — the 4.7 memory-wiki / memory-core / memory-lancedb refactor didn't propagate into the active-memory plugin path

Happy to run additional diagnostics or share full log chunks if helpful.