[Bug] 2026.4.2 升级导致数据层不兼容，引发 Agent 无限循环

[Clivilwalker]

问题描述

从 2026.3.31 升级到 2026.4.2 后，Agent 进入无限循环重启状态。即使回退到 2026.3.31、更换 smart extraction 模型、关闭 auto-recall，问题依然存在。最终只能通过服务器备份恢复解决。

复现步骤

1. 环境：OpenClaw 2026.3.31，memory-lancedb-pro@1.1.0-beta.10，数据文件正常
2. 升级到 2026.4.2
3. 启动 Gateway，与 Agent 对话
4. Agent 开始循环重启

关键日志

// 循环触发点
"memory-lancedb-pro: auto-recall timed out after 5000ms; skipping memory injection"

// 即使配置已修改，仍尝试 auto-recall
// 配置：smartExtraction: ON → OFF, model: k2p5 → MiniMax

尝试过的修复（均失败）

操作	结果
更换 smart extraction 模型为 MiniMax	❌ 仍循环
关闭 auto-recall	❌ 仍循环（日志显示仍在尝试）
回退到 2026.3.31	❌ 仍循环
服务器备份恢复	✅ 正常

根因分析

疑似 2026.4.2 修改了 LanceDB 的存储格式或索引结构，导致：

1. 升级后数据文件处于异常状态
2. 回退版本后，3.31 无法正确解析新格式
3. auto-recall 功能可能硬编码触发，配置关闭无效

建议修复

1. 增加数据迁移机制（2026.3.x → 2026.4.x）
2. 或版本隔离（不同版本使用独立数据目录）
3. 确保 auto-recall 配置项可完全禁用

环境信息

• OpenClaw: 2026.4.2 / 2026.3.31
• memory-lancedb-pro: 1.1.0-beta.10
• Node.js: v22.22.0
• OS: Linux VM-0-12-opencloudos

[AliceLJY]

Thanks for the detailed report — this is a serious issue.

A few observations:

1. The "config off but still fires" symptom is key.
If auto-recall: off in config still triggers auto-recall, that points to either:

• Config not being re-parsed after change (cached in memory, needs gateway restart)
• A hook registration that persists beyond config toggle — before_prompt_build hooks are registered at init time and may not respect runtime config changes

2. Relationship to other 4.2 issues
This may be related to:

• memory-lancedb-pro v1.0.32 incompatible with OpenClaw 4.2 (stringEnum not a function) #487 (stringEnum not a function) — also triggered by 4.2 upgrade
• 0401版本Openclaw安装这个插件以后网关一直不停重启 #465 (gateway restart loop) — similar infinite-restart pattern
• Memory lost on gateway restart: agent_end hook never fires #435 (memory lost on gateway restart) — agent_end hook never fires

3. The "backup restore fixes it" clue
This strongly suggests the LanceDB data files were corrupted or left in an incompatible state by the 4.2 upgrade. Rolling back the OpenClaw version alone doesn't help because the data has already been touched.

Recommendations for affected users:

1. Immediate workaround: If you still have the backup, restore it and stay on 3.31 until we ship a fix
2. If no backup: Try deleting the LanceDB data directory (data/lancedb/ or the configured path) — you'll lose stored memories but the plugin should reinitialize cleanly

For us to investigate:

• Does 4.2 change the LanceDB table schema or index format?
• Is there a migration path we need to add for existing data?
• The auto-recall config bypass needs a separate fix — config changes should take effect without a full data wipe

Could you share the full startup log (with DEBUG=memory-lancedb-pro* if possible)? That would help pinpoint exactly where the loop starts.

[AliceLJY]

OpenClaw 4.2 data layer incompatibility — resolved in subsequent releases. Please upgrade to latest plugin and OpenClaw versions.