memory-lancedb-pro @ master (post v1.1.0-beta.10): agent_end hook blocked under external (non-bundled) install — data stays in-memory, never flushes to memories.lance

[OKmiao88888888]

Bug Description

After a clean install of memory-lancedb-pro as an external (non-bundled) plugin under ~/.openclaw/workspace/plugins/memory-lancedb-pro, the agent_end typed hook is silently blocked at runtime. This causes smartExtraction and memoryReflection to never fire, and crucially, the in-memory autoCapturePendingIngressTexts queue never gets flushed to memories.lance — data accumulated in the current Gateway process is at risk of total loss on restart.

Environment

• OpenClaw: 2026.6.1-beta.2
• memory-lancedb-pro: master @ commit 15b6529 (5 commits past the latest GitHub release v1.1.0-beta.10, 2026-03-23)
  • Fix reflection legacy upgrade classification #846 fix reflection legacy upgrade classification
  • Fix default scopes for memory management tools #844 fix management tool default scopes
  • Warn on costly auto-recall rerank config #843 Warn on costly auto-recall rerank config
  • fix: clarify plugin config preflight errors #839 fix: clarify plugin config preflight errors
  • feat: add Obsidian memory sync command #849 feat: add Obsidian memory sync command
• v1.1.0-beta.10 (the latest GitHub release) is also affected — the behavior is not specific to any commit
• Plugin install path: ~/.openclaw/workspace/plugins/memory-lancedb-pro (external, non-bundled)
• OS: Windows 10

Reproduction Steps

1. Install memory-lancedb-pro as an external plugin in OpenClaw (not bundled)
2. Enable smartExtraction: true and sessionStrategy: "memoryReflection" in plugin config
3. Run the agent for several conversation turns
4. Check ~/.openclaw/memory/lancedb-pro/memories.lance mtime — it does not update after agent_end
5. Check ~/.openclaw/memory/lancedb-pro/.memory-write.lock — last modified at plugin install time, 0 bytes, no active write lock

Expected Behavior

agent_end should trigger smartExtraction to write structured facts to LanceDB, and memoryReflection should store session summaries. The autoCapturePendingIngressTexts queue should be drained on agent_end and the entries persisted to memories.lance.

Actual Behavior

agent_end hook is silently dropped at the registry level. No error shown to the user. smartExtraction and memoryReflection never run. autoCapturePendingIngressTexts queue grows without bound in process memory. memories.lance mtime stays at the original install timestamp.

Workaround (unverified by maintainers)

Add the following to openclaw.json under plugins.entries.memory-lancedb-pro (note: hooks is a sibling of config, not nested inside it):

"hooks": {
  "allowConversationAccess": true
}

This matches the suggested workaround in #707. However, this workaround has not yet been confirmed by a maintainer, and a restart of the Gateway may be required for the change to take effect (unverified).

Related

• [BUG] agent_end hook blocked in OpenClaw — smartExtraction and memoryReflection silently fail #707 — original report of the same agent_end hook blocking behavior. Currently open with no maintainer response. Workaround is reporter-supplied, not officially endorsed.
• Memory lost on gateway restart: agent_end hook never fires #435 — separate report of memory loss on gateway restart (possibly related, but different root cause)

Request

Could a maintainer please:

1. Confirm whether hooks.allowConversationAccess: true is the correct user-side config, or if a different option name / path is expected
2. Consider adding a clearer error at plugin startup when typed hooks are blocked (per the priority suggestions in [BUG] agent_end hook blocked in OpenClaw — smartExtraction and memoryReflection silently fail #707)
3. Update the README to document the required hooks.allowConversationAccess opt-in for external (non-bundled) installs

[TurboTheTurtle]

Rechecked this against current master, and the source-side fix for the hook-blocking part appears to be present now.

Evidence on master:

• openclaw.plugin.json declares hooks.allowConversationAccess: true, so external/non-bundled installs no longer need every user to add the workaround manually in openclaw.json.
• the manifest declares contracts.tools for the memory/self-improvement tool surface.
• related manifest compatibility landed across fix: declare contracts.tools for OpenClaw 2026.5.x compatibility #742 / fix: add OpenClaw 2026.5 compiled runtime compatibility #755 / fix: declare memory_reflection_resolve contract and remove duplicate contracts key #869, and test/plugin-manifest-regression.mjs covers the hook permission and tool contract declarations.

This overlaps with #707. I would treat the original “typed agent_end hook is blocked for external installs” part as covered on current master. The remaining caveat is distribution: users on the last published/released package may still see this until a fresh release/package publish includes the updated manifest. If agent_end fires but extraction still yields zero texts, that is likely the narrower follow-up path described in #707 rather than this permission-blocking issue.