Possible regression with OpenClaw 2026.3.28: sessionStrategy resolves to none and memory-reflection hooks are disabled

[jlin53882]

Summary

After updating OpenClaw from a working older version to 2026.3.28, memory-lancedb-pro behavior changed unexpectedly in our environment:

• sessionStrategy appears to resolve to none
• memory-reflection hooks are disabled at startup
• downstream behavior becomes inconsistent / degraded
• rolling OpenClaw back to 2026.3.24 restores expected behavior immediately

At this point I cannot determine whether the regression is in OpenClaw 2026.3.28 itself, or whether memory-lancedb-pro is relying on older config/runtime assumptions that 2026.3.28 changed.

Observed behavior

With OpenClaw 2026.3.28, gateway logs show:

session-strategy: using none (plugin memory-reflection hooks disabled)

Later, after forcing:

"plugins": {
  "entries": {
    "memory-lancedb-pro": {
      "config": {
        "sessionStrategy": "memoryReflection"
      }
    }
  }
}

and restarting, logs show:

memory-reflection: integrated hooks registered (command:new, command:reset, after_tool_call, before_prompt_build, session_end)

So the plugin behavior definitely changes depending on sessionStrategy resolution.

Why this looks like a regression

• Before the OpenClaw 2026.3.28 update, our setup worked normally.
• After updating to 2026.3.28, the plugin started behaving as if sessionStrategy were none.
• Rolling back OpenClaw to 2026.3.24 makes the problem disappear.

That strongly suggests a compatibility break or config-resolution behavior change across OpenClaw versions.

Environment

• OS: Windows 10.0.26200 x64
• Node: 24.13.0
• OpenClaw: 2026.3.28 (problematic)
• OpenClaw rollback target: 2026.3.24 (works)
• memory-lancedb-pro: 1.1.0-beta.10
• Plugin DB path: ~/.openclaw/memory/lancedb-pro-jina1024

Relevant plugin config

{
  "autoCapture": true,
  "autoRecall": true,
  "retrieval": {
    "mode": "hybrid",
    "rerank": "cross-encoder",
    "rerankProvider": "siliconflow",
    "rerankModel": "BAAI/bge-reranker-base",
    "rerankEndpoint": "http://127.0.0.1:18799/v1/rerank",
    "rerankApiKey": "local"
  },
  "embedding": {
    "baseURL": "http://localhost:11434/v1",
    "model": "jina-v5-retrieval-test",
    "dimensions": 1024
  },
  "autoRecallTimeoutMs": 120000
}

Additional notes

I also observed degraded runtime behavior in 2026.3.28 such as:

[EventQueue] Slow listener detected: InteractionEventListener took 4254ms for event INTERACTION_CREATE

and repeated websocket fallback logs like:

[ws-stream] WebSocket connect failed ... Unexpected server response: 500

Those may be OpenClaw-side regressions rather than memory-lancedb-pro bugs, but I am including them because they appeared at the same time as the plugin behavior change.

Request

Could you help determine:

1. whether memory-lancedb-pro@1.1.0-beta.10 assumes older OpenClaw behavior around sessionStrategy
2. whether OpenClaw 2026.3.28 changed plugin config resolution in a way that now forces / defaults this plugin into none
3. whether memory-lancedb-pro should explicitly guard / normalize sessionStrategy instead of depending on upstream resolution

If needed, I can provide more logs.

[AliceLJY]

Linking with #396 — same environment, same version boundary.

The fact that explicitly setting sessionStrategy in plugin config fixes it suggests OpenClaw 3.28 changed how plugin config defaults are resolved. This is upstream behavior — suggest filing on the OpenClaw repo so the maintainers can investigate the config resolution change between 3.24 and 3.28.

[jlin53882]

@AliceLJY I have already reported this to OpenClaw and asked them to investigate the cause.
openclaw/openclaw#56881

[jlin53882]

Thanks for the report! We experienced the same issue.

Root cause identified: On session start (reset/new), the OpenClaw built-in memory-search loads nomic-embed-text while memory-lancedb-pro loads jina-v5-retrieval-test simultaneously. These two embedding models compete for Ollama's 8 CPU threads, causing embedding latency to degrade from ~100ms to 1-5+ seconds.

Workaround applied: Changed OpenClaw built-in default embedding model to jina-v5-retrieval-test, so only one embedding model is loaded at a time.

This has been reported to OpenClaw as well: openclaw/openclaw#57149

Note: autoRecallTimeoutMs has a hard limit of 120000ms in memory-lancedb-pro and cannot be raised higher.