openclaw - 💡(How to fix) Fix [bug] active-memory plugin: 7-15s framework overhead per turn even when model responds in <2s [1 comments, 2 participants]

Environment

• OpenClaw 2026.4.27
• Docker deployment, Linux ARM64 (Apple Silicon host)
• Anthropic API key working (verified: direct curl returns in ~1s)

Summary

The active-memory plugin's before_prompt_build hook adds 7-15s of latency per chat turn, even though the underlying model (Anthropic Claude Haiku 4-5) responds within 2s. This blocks every chat reply on the configured agent and makes it unusable in interactive contexts.

Repro

Single agent (main) with active-memory enabled, defaults:

"active-memory": {
  "enabled": true,
  "config": {
    "agents": ["main"],
    "allowedChatTypes": ["direct"],
    "queryMode": "recent",
    "promptStyle": "balanced",
    "model": "anthropic/claude-haiku-4-5",
    "modelFallback": "openai/gpt-4o-mini",
    "timeoutMs": 10000,
    "maxSummaryChars": 300,
    "thinking": "off"
  }
}

Send any direct chat message to the configured agent.

Observed (gateway logs)

active-memory: agent=main start timeoutMs=15000 queryChars=369
[hooks] before_prompt_build handler from active-memory failed: timed out after 15000ms
active-memory: ... done status=timeout elapsedMs=15979 summaryChars=0
[events] Cost: claude/claude-haiku-4-5 [main] — 2967 in, 0cr, 0cw, 181 out [active-memory]

The cost log proves the model did respond with 181 tokens during the call, but the framework timed out before delivering the result. A subsequent identical query: status=empty elapsedMs=7107 — still 7s, even with cacheTtlMs: 60000 configured.

Liveness diagnostic during the call:

liveness warning: eventLoopDelayMaxMs=8157.9 eventLoopUtilization=0.842 cpuCoreRatio=0.603

— event loop blocked ~8s by the embedded run.

Tested config variations (none helped)

• Models: anthropic/claude-haiku-4-5, openai-codex/gpt-5.5, openai/gpt-4o-mini — all hit timeout/empty
• Settings: queryMode: "message", promptStyle: "preference-only", recentUserTurns: 0, recentAssistantTurns: 0, cacheTtlMs: 60000, thinking: "off" — no measurable improvement
• Memory backend: Vector: ready and FTS: ready (after side-issue fix below)

Side issue: missing dep sqlite-vec

sqlite-vec is imported by dist/engine-storage-*.js but not declared in plugin-runtime-deps/*/package.json. Out of the box: Vector: unavailable. Manual npm install sqlite-vec sqlite-vec-linux-arm64 in the runtime-deps directory was needed before vector search worked.

sqlite-vec unavailable: Cannot find package 'sqlite-vec' imported from
  /home/node/.openclaw/plugin-runtime-deps/openclaw-2026.4.27-*/dist/engine-storage-CO3qFXo7.js
Did you mean to import "sqlite-vec/index.cjs"?

Suggest declaring sqlite-vec (and platform-specific binaries) as bundled runtime dependencies.

Expected

Active-memory before_prompt_build hook completes in <3s when the underlying model returns in <2s. If framework overhead is intrinsic, the docs should state a realistic minimum timeoutMs (15s currently insufficient).

Workaround

For now: enabled: false. Will re-enable once a fix lands.

Happy to share more logs / try patches if useful.