QMD memory for OpenClaw: what still works in Lobsterland

QMD memory is the Query Markup Documents search engine. OpenClaw removed it in version 2026.8.1, released August 31, 2026, and replaced it with built-in Memory. Lobsterland still supports QMD as a managed compatibility path — the platform installs a pinned QMD package, emits the supported backend contract, and restarts the instance — but only on catalog images older than 2026.8.1.

What is QMD memory?

QMD stands for Query Markup Documents. It is an on-device search engine for Markdown notes, documentation, transcripts, and other local knowledge. The QMD package offers three distinct retrieval commands: keyword search, vector search, and hybrid query plus reranking.

Package capability is not the same as the mode a host selects. Lobsterland currently selects search, the fast BM25 keyword path, for agent-memory retrieval. The managed pods also keep a local embedding model available and schedule embedding maintenance, but the default retrieval mode is not QMD's hybrid query path.

Is QMD still part of OpenClaw?

No. OpenClaw retired QMD in version 2026.8.1, released August 31, 2026, and made built-in Memory the only engine. That release also removed the legacy agents.defaults.memorySearch configuration schema. Upstream’s documented migration is to run openclaw doctor --fix, which strips the retired QMD settings, preserves compatible rows already in the agent database, and rebuilds the index from canonical Markdown — and which retires QMD-only reranking, query expansion, and cross-agent transcript search with no replacement. Old guides that tell you to set memory.backend = "qmd" describe an OpenClaw contract that ended at 2026.7.x.

Lobsterland’s QMD toggle is a managed compatibility surface for the pre-2026.8.1 images the platform still supports. It is not a promise that the same block works on a current self-hosted checkout, and it is not offered on 2026.8.1 instances at all.

How does managed QMD work in Lobsterland?

Enable QMD from a running instance's Settings view. Lobsterland saves the setting, installs its pinned QMD package and native SQLite dependency in the pod, writes the managed config, and restarts the instance. There is no package command or config file for the customer to maintain.

As of August 26, 2026, Lobsterland generates this JSON contract:

{
  "memory": {
    "backend": "qmd",
    "citations": "auto",
    "qmd": {
      "searchMode": "search",
      "includeDefaultMemory": true,
      "update": {
        "onBoot": false,
        "waitForBootSync": false,
        "interval": "15m",
        "debounceMs": 15000,
        "embedInterval": "12h",
        "embedTimeoutMs": 3600000
      },
      "limits": { "maxResults": 6, "timeoutMs": 4000 },
      "scope": {
        "default": "deny",
        "rules": [{ "action": "allow", "match": { "chatType": "direct" } }]
      }
    }
  }
}

Which QMD search mode does Lobsterland use?

The distinction matters. QMD is a hybrid-capable package; Lobsterland's current OpenClaw memory path deliberately uses its keyword mode. Collapsing those two facts over-promises what an individual managed lookup does.

How do you enable QMD in managed OpenClaw hosting?

These steps apply only to instances running an OpenClaw image older than 2026.8.1. On 2026.8.1 and newer there is no QMD toggle to turn on: Settings shows a Built-in Memory card instead, and durable Markdown memories are searchable without a separate plugin.

1. Start the OpenClaw instance.
2. Open Settings.
3. Turn on Enable QMD memory backend.
4. Wait for the instance to finish reprovisioning and return to Running.
5. Ask for a distinctive phrase from MEMORY.md and check that the answer carries a source path.

The setting requires a running instance because Lobsterland installs QMD into the live pod before restarting it. If you want the host to own that installation and version pin, use managed OpenClaw hosting with pinned runtime dependencies.

What does QMD index on a managed instance?

includeDefaultMemory: true points QMD at OpenClaw's default workspace memory material, including MEMORY.md and the memory directory. You can inspect and edit those files through Lobsterland's workspace management tools without maintaining a separate QMD collection by hand.

QMD retrieval is direct-chat-only under the managed default. That is a retrieval boundary, not a claim that group conversations are never stored anywhere. If you need the underlying workspace model first, read the guide to OpenClaw memory files.

What are the current QMD limits?

• OpenClaw 2026.8.1 and newer no longer expose QMD as a memory backend at all.
• Lobsterland’s compatibility path is tied to managed OpenClaw images older than 2026.8.1.
• Managed retrieval uses BM25 keyword search by default, not QMD's hybrid query mode.
• A 15-minute background scan means a new memory is not guaranteed to be searchable immediately.
• QMD results are allowed in direct chats by default; other chat types are denied by the emitted scope rule.
• QMD handles durable file retrieval. It does not replace conversation compaction. If you need both layers, see how Lossless Claw and QMD fit together.

QMD memory FAQ