cant get extra path for qmd memory to work | Friends of the Crustacean 🦞🤝 | Page 1

true hound Apr 13, 2026, 3:20 AM

#

it continue to default to the default memory not the extra path im adding. I have been on the latest stable version.

OpenClaw 2026.4.11 (769908e)

ripe ruinBOT Apr 13, 2026, 3:20 AM

#

ionic juncoBOT Apr 13, 2026, 3:20 AM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

meager moonBOT Apr 13, 2026, 3:22 AM

#

Let me verify the docs and the actual config key in source, so I don’t guess.

#

Yep, I checked the docs + source, and this is probably the gotcha:

extraPaths does not replace the default memory. It only adds extra indexed search roots.

So with QMD:

agents.defaults.memorySearch.extraPaths = additive
memory.qmd.paths = additive
memory.qmd.includeDefaultMemory defaults to true

That means OpenClaw will still index MEMORY.md + memory/**/*.md unless you disable it.

If you want only your extra path for QMD indexing, use something like:

#

  memory: {
    backend: "qmd",
    qmd: {
      includeDefaultMemory: false,
      paths: [
        { name: "notes", path: "/your/extra/path", pattern: "**/*.md" }
      ]
    }
  }
}

Two other important gotchas:

Bootstrap/system prompt still prefers MEMORY.md
Extra QMD paths are for search/indexing, not automatic prompt injection.

#

Docs explicitly say bootstrap auto-loads MEMORY.md, not arbitrary extra files/dirs.

Writes/promotions still go to MEMORY.md
Long-term promotion is still centered on MEMORY.md, not an arbitrary extra path.

Quick check to confirm what OpenClaw is actually seeing:

openclaw memory status --deep --index --verbose

That should show whether your extra path is in the memory index.

Docs:

#

If you paste your current memory + agents.defaults.memorySearch config block, I can sanity-check the exact shape.

true hound Apr 13, 2026, 3:24 AM

#

i dont only want the extra path. its just not even using the extra path

meager moonBOT Apr 13, 2026, 3:25 AM

#

Got it, then I’d check these 4 things first, because one of them is usually the culprit:

Group/chat scope
If you’re testing in Discord channels/threads, QMD is DM-only by default.
So the extra path can be indexed fine, but results won’t be surfaced there unless you allow groups/channels.
```
{
  memory: {
    backend: "qmd",
    qmd: {
      scope: {
        default: "allow"
      }
    }
  }
```

#

extraPaths only indexes Markdown
agents.defaults.memorySearch.extraPaths defaults to **/*.md.
If your extra path has non-.md files, it won’t pick them up.

For more control, use memory.qmd.paths:
```
{
  memory: {
    backend: "qmd",
    qmd: {
      paths: [
```

#

           { name: "extra", path: "/your/path", pattern: "**/*.md" }
         ]
       }
     }
   }

It may not be indexed yet
Force a refresh:

openclaw memory status --deep --index --verbose

Test the index directly, not the assistant behavior
Use a phrase that exists only in a file under the extra path:

#

   openclaw memory search --query "some unique phrase from that file" --json

If that search returns nothing, it’s an indexing/config issue.
If it returns the file, then QMD is working and the issue is retrieval/scope.

If you paste:

your memory block
your agents.defaults.memorySearch block
output of openclaw memory status --deep --index --verbose

I can tell you exactly what’s off.

true hound Apr 13, 2026, 3:35 AM

#

so testing the index directly is what ive been doing this entire time.
i have to break the output into multiple messages for length so wait until i say "im done" for you to reply.

totaluser@Total-Hardware:~$ openclaw memory status --deep --index --verbose
20:30:54-07:00 [plugins] plugins.allow is empty; discovered non-bundled plugins may auto-load: lossless-claw (/home/totaluser/.openclaw/extensions/lossless-claw/dist/index.js). Set plugins.allow to explicit trusted ids.

🦞 OpenClaw 2026.4.11 (769908e)
Your task has been queued; your dignity has been deprecated.

20:30:59-07:00 [bedrock-mantle-discovery] Mantle IAM token generation unavailable
20:31:06-07:00 [memory] qmd collection add skipped for memory-alt-main: qmd collection add /home/totaluser/.openclaw/workspace --name memory-alt-main --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-main (qmd://memory-root-main/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-main'

│
◇
20:31:19-07:00 [memory] qmd collection add skipped for memory-alt-forge: qmd collection add /home/totaluser/.openclaw/forge --name memory-alt-forge --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-forge (qmd://memory-root-forge/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-forge'

│
◇
20:31:33-07:00 [memory] qmd collection add skipped for memory-alt-scout: qmd collection add /home/totaluser/.openclaw/scout --name memory-alt-scout --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-scout (qmd://memory-root-scout/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-scout'

│
◇

#

20:31:45-07:00 [memory] qmd collection add skipped for memory-alt-vital: qmd collection add /home/totaluser/.openclaw/vital --name memory-alt-vital --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-vital (qmd://memory-root-vital/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-vital'

│
◇
20:31:58-07:00 [memory] qmd collection add skipped for memory-alt-sentinel: qmd collection add /home/totaluser/.openclaw/sentinel --name memory-alt-sentinel --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-sentinel (qmd://memory-root-sentinel/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-sentinel'

│
◇
20:32:18-07:00 [memory] qmd collection add skipped for memory-alt-alfred: qmd collection add /home/totaluser/.openclaw/alfred --name memory-alt-alfred --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-alfred (qmd://memory-root-alfred/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-alfred'

│
◇
20:32:23-07:00 [memory] qmd collection add skipped for memory-alt-rush: qmd collection add /home/totaluser/.openclaw/rush --name memory-alt-rush --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-rush (qmd://memory-root-rush/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-rush'

│
◇
20:32:28-07:00 [memory] qmd collection add skipped for memory-alt-cipher: qmd collection add /home/totaluser/.openclaw/cipher --name memory-alt-cipher --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-cipher (qmd://memory-root-cipher/)
Pattern: **/*.md

#

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-cipher'

│
◇
Memory index complete.
Memory Search (main)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 1529/165 files · 1529 chunks
Dirty: no
Store: ~/.openclaw/agents/main/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/workspace

Dreaming: off
Embeddings: ready
By source:
memory · 1321/63 files · 1321 chunks
sessions · 208/102 files · 208 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 1 entries · 0 promoted · 1 concept-tagged · 0 spaced · scripts=1 latin
Recall path: ~/.openclaw/workspace/memory/.dreams/short-term-recall.json
Recall updated: 2026-04-12T23:47:20.338Z
QMD audit: ~/.openclaw/agents/main/qmd/xdg-cache/qmd/index.sqlite · 59420672 bytes · 5 collections

Memory index complete.
Memory Search (forge)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 815/185 files · 815 chunks
Dirty: no
Store: ~/.openclaw/agents/forge/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/forge
Dreaming: off
Embeddings: ready
By source:
memory · 581/49 files · 581 chunks
sessions · 234/136 files · 234 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 4 entries · 0 promoted · 4 concept-tagged · 0 spaced · scripts=4 latin
Recall path: ~/.openclaw/forge/memory/.dreams/short-term-recall.json
Recall updated: 2026-04-13T02:31:31.898Z
QMD audit: ~/.openclaw/agents/forge/qmd/xdg-cache/qmd/index.sqlite · 43405312 bytes · 5 collections

Memory index complete.
Memory Search (scout)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 838/205 files · 838 chunks
Dirty: no
Store: ~/.openclaw/agents/scout/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/scout
Dreaming: off
Embeddings: ready
By source:
memory · 649/93 files · 649 chunks
sessions · 189/112 files · 189 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 0 entries · 0 promoted · 0

#

concept-tagged · 0 spaced
Recall path: ~/.openclaw/scout/memory/.dreams/short-term-recall.json
QMD audit: ~/.openclaw/agents/scout/qmd/xdg-cache/qmd/index.sqlite · 39432192 bytes · 5 collections
Memory index complete.
Memory Search (vital)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 1112/175 files · 1112 chunks
Dirty: no
Store: ~/.openclaw/agents/vital/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/vital
Dreaming: off
Embeddings: ready
By source:
memory · 899/35 files · 899 chunks
sessions · 213/140 files · 213 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 0 entries · 0 promoted · 0 concept-tagged · 0 spaced
Recall path: ~/.openclaw/vital/memory/.dreams/short-term-recall.json
QMD audit: ~/.openclaw/agents/vital/qmd/xdg-cache/qmd/index.sqlite · 68493312 bytes · 5 collections
Memory index complete.
Memory Search (sentinel)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 1183/461 files · 1183 chunks
Dirty: no
Store: ~/.openclaw/agents/sentinel/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/sentinel
Dreaming: off
Embeddings: ready
By source:
memory · 754/166 files · 754 chunks
sessions · 429/295 files · 429 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 0 entries · 0 promoted · 0 concept-tagged · 0 spaced
Recall path: ~/.openclaw/sentinel/memory/.dreams/short-term-recall.json
QMD audit: ~/.openclaw/agents/sentinel/qmd/xdg-cache/qmd/index.sqlite · 37052416 bytes · 5 collections

Memory index complete.
Memory Search (alfred)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 489/42 files · 489 chunks
Dirty: no
Store: ~/.openclaw/agents/alfred/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/alfred
Dreaming: off
Embeddings: ready
By source
memory · 434/12 files · 434 chunks
sessions · 55/30 files · 55 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 0 entries · 0 promoted · 0 concept-tagged · 0 spaced

#

Recall path: ~/.openclaw/alfred/memory/.dreams/short-term-recall.json
QMD audit: ~/.openclaw/agents/alfred/qmd/xdg-cache/qmd/index.sqlite · 24944640 bytes · 5 collections

Memory index complete.
Memory Search (rush)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 158/8 files · 158 chunks
Dirty: no
Store: ~/.openclaw/agents/rush/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/rush
Dreaming: off
Embeddings: ready
By source:
memory · 151/5 files · 151 chunks
sessions · 7/3 files · 7 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 4 entries · 0 promoted · 4 concept-tagged · 0 spaced · scripts=3 latin, 1 other
Recall path: ~/.openclaw/rush/memory/.dreams/short-term-recall.json
Recall updated: 2026-04-13T01:09:48.567Z
QMD audit: ~/.openclaw/agents/rush/qmd/xdg-cache/qmd/index.sqlite · 6615040 bytes · 5 collections

Memory index complete.
Memory Search (cipher)
Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 154/77 files · 154 chunks
Dirty: no
Store: ~/.openclaw/agents/cipher/qmd/xdg-cache/qmd/index.sqlite
Workspace: ~/.openclaw/cipher
Dreaming: off
Embeddings: ready
By source:
memory · 28/10 files · 28 chunks
sessions · 126/67 files · 126 chunks
Vector: ready
Batch: disabled (failures 0/0)
Recall store: 0 entries · 0 promoted · 0 concept-tagged · 0 spaced
Recall path: ~/.openclaw/cipher/memory/.dreams/short-term-recall.json
QMD audit: ~/.openclaw/agents/cipher/qmd/xdg-cache/qmd/index.sqlite · 6086656 bytes · 5 collections

#

im done

meager moonBOT Apr 13, 2026, 3:38 AM

#

Thanks, this is actually useful.

From this output alone, QMD does not look broken. A few things stand out:

Provider: qmd, Embeddings: ready, Dirty: no
each agent has QMD audit ... 5 collections
memory-alt-* skipped is happening because an existing memory-root-* collection already covers that path with **/*.md

So my read is:

your extra path is probably being indexed
the ugly part is more likely stale/legacy QMD collection state
memory status still does not tell us which exact collection/path won the query

The suspicious bit is this:

#

memory-alt-main ... failed ...
existing collection: memory-root-main
Pattern: **/*.md

That suggests you already have an older broad workspace-root collection hanging around, which can make this really confusing.

What I need next is not more memory status. I need these 3 things:

openclaw config get memory --json
openclaw config get agents.defaults.memorySearch --json
openclaw memory search --agent main --query "EXACT UNIQUE TEXT THAT ONLY EXISTS IN THE EXTRA PATH" --json

If you’re not testing main, swap in the actual agent id.

#

Also, one key detail: is your extra path

outside the workspace, or
inside ~/.openclaw/workspace?

Because if it’s inside the workspace, that old broad memory-root-* collection may already be swallowing it into the default root collection, which makes it look like the extra path is being ignored when it actually isn’t.

So at this point I’m leaning toward stale QMD collection layout / overlap, not “extraPaths does nothing”.

Paste those 3 outputs and I can narrow it down fast.

true hound Apr 13, 2026, 3:40 AM

#

like if there is some text like "quantum mechanics" in one of the files of in the extra path it should be able to find it if its working right?

meager moonBOT Apr 13, 2026, 3:40 AM

#

Yes, for a direct CLI search, that’s the right expectation.

If:

the file is actually indexed,
it’s a supported file type, and
"quantum mechanics" is a reasonably unique phrase in that extra path,

then:

openclaw memory search --agent main --query "quantum mechanics" --json

should return a hit from that file.

A couple of caveats:

#

In normal chat, the assistant might not call memory search every time, so chat behavior is fuzzier.
But with openclaw memory search ..., it should be much more direct.
If the phrase exists only in the extra path and this command still returns nothing, then yeah, I’d treat that as extra path indexing not working correctly or stale QMD collection state.

Best test is an even more unique string, like a full sentence or weird phrase that appears in exactly one file. If that still misses, that’s a real signal.

true hound Apr 13, 2026, 3:45 AM

#

there are two halves

totaluser@Total-Hardware:~$ openclaw config get memory --json
openclaw config get agents.defaults.memorySearch --json
openclaw memory search --agent main --query "quantum mechanics" --json
{
"backend": "qmd",
"qmd": {
"command": "/home/totaluser/.local/share/pnpm/qmd",
"paths": [
{
"path": "/home/totaluser/.openclaw/yr-knowledge",
"name": "yr-knowledge",
"pattern": "**/.md"
}
],
"sessions": {
"enabled": true
},
"scope": {
"default": "deny",
"rules": [
{
"action": "allow",
"match": {
"chatType": "direct"
}
},
{
"action": "allow",
"match": {
"chatType": "channel"
}
}
]
}
}
}
{
"enabled": true,
"experimental": {
"sessionMemory": true
}
}
20:42:09-07:00 [plugins] plugins.allow is empty; discovered non-bundled plugins may auto-load: lossless-claw (/home/totaluser/.openclaw/extensions/lossless-claw/dist/index.js). Set plugins.allow to explicit trusted ids.
20:42:22-07:00 [memory] qmd collection add skipped for memory-alt-main: qmd collection add /home/totaluser/.openclaw/workspace --name memory-alt-main --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-main (qmd://memory-root-main/)
Pattern: **/.md
Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-main'
20:42:23-07:00 [memory] qmd search failed because a managed collection is missing; repairing collections and retrying once
20:42:27-07:00 [memory] qmd collection add skipped for memory-alt-main: qmd collection add /home/totaluser/.openclaw/workspace --name memory-alt-main --glob memory.md failed (code 1): A collection already exists for this path and pattern:

this is part 1

#

20:42:28-07:00 [memory] qmd search failed: Error: qmd search quantum mechanics --json -n 6 -c memory-alt-main failed (code 1): Collection not found: memory-alt-main
20:42:28-07:00 [memory] qmd memory failed; switching to builtin index: qmd search quantum mechanics --json -n 6 -c memory-alt-main failed (code 1): Collection not found: memory-alt-main
{
"results": [
{
"id": "eda31c292358abf1f5b160b92199894d84329bf4feb9c213ebe30aa849fa0e55",
"path": "memory/digest-2026-04-03.md",
"startLine": 1,
"endLine": 31,
"score": 0.8343585020974064,
"textScore": 0.8343585020974064,
"snippet": "Scout Daily Digest — 2026-04-03\n\nFrom: Scout\nTo: Total\nTime: 13:06 PST\n\n---\n\n 🔴 High Priority\n\n Health Privacy Toggle — Technical Deep Dive\nResearch file: workspace/research/2026-04-03-health-privacy-toggle-mechanism.md\nRouted to: Vital's HEALTH_INBOX\n\nWhat I found:\nThe Health Privacy toggle controls per-data-type permissions in Apple HealthKit (iOS) or Google Health Connect (Android) — not a single global switch. \n\nApple HealthKit mechanics:\n- Toggling off sets HKAuthorizationStatus.sharingDenied per data type — writes fail with HKError.healthDataUnavailable, reads return empty (indistinguishable from no data existing)\n- User controls at: Settings >",
"source": "memory"
}
]
}
20:42:34-07:00 [memory] chunks written for MEMORY.md without vector embeddings — chunks_vec not updated (sqlite-vec unavailable). Vector recall degraded for this file.
there are a ton more of these chunks written for MEMORY.md messages.

meager moonBOT Apr 13, 2026, 3:47 AM

#

Yep, this is the smoking gun:

qmd search ... -c memory-alt-main failed: Collection not found: memory-alt-main
qmd memory failed; switching to builtin index

So the problem is not your memory.qmd.paths config. That part looks correct.

What’s happening is:

OpenClaw tries to use QMD
QMD’s managed collection state is broken/stale
OpenClaw falls back to the builtin index
you get results from default memory instead of your extra QMD path

And this line:

#

chunks written for MEMORY.md without vector embeddings ...

is basically fallout from the builtin fallback, not the root cause.

What I’d do next

First, check your QMD version:

qmd --version

Then, for the agent you’re testing (main here), do a clean QMD reset with a backup:

#

openclaw gateway stop
mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
openclaw gateway start
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "quantum mechanics" --json

Why this helps:

QMD collections are persisted inside that per-agent QMD state
your logs show that state is inconsistent (memory-alt-main is expected but missing)
moving the dir aside forces OpenClaw to rebuild the managed QMD collections from config

If main starts working after that, you can repeat it for any other affected agents.

If it still errors after a clean reset, then I’d call it a real bug and ask you to file it here:
https://github.com/openclaw/openclaw/issues

#

If you want, paste the output of qmd --version first, because I’m a little suspicious this may also involve an older QMD CLI behavior.

true hound Apr 13, 2026, 3:49 AM

#

totaluser@Total-Hardware:~$ qmd --version
Usage:
qmd collection add [path] --name <name> --mask <pattern> - Create/index collection
qmd collection list - List all collections with details
qmd collection remove <name> - Remove a collection by name
qmd collection rename <old> <new> - Rename a collection
qmd ls [collection[/path]] - List collections or files in a collection
qmd context add [path] "text" - Add context for path (defaults to current dir)
qmd context list - List all contexts
qmd context rm <path> - Remove context
qmd get <file>[:line] [-l N] [--from N] - Get document (optionally from line, max N lines)
qmd multi-get <pattern> [-l N] [--max-bytes N] - Get multiple docs by glob or comma-separated list
qmd status - Show index

two halves this is part 1

#

qmd embed [-f] - Create vector embeddings (800 tokens/chunk, 15% overlap)
qmd cleanup - Remove cache and orphaned data, vacuum DB
qmd search <query> - Full-text search (BM25)
qmd vsearch <query> - Vector similarity search
qmd query <query> - Combined search with query expansion + reranking
qmd mcp - Start MCP server (stdio transport)
qmd mcp --http [--port N] - Start MCP server (HTTP transport, default port 8181)
qmd mcp --http --daemon - Start MCP server as background daemon
qmd mcp stop - Stop background MCP daemon

Global options:
--index <name> - Use custom index name (default: index)

Search options:
-n <num> - Number of results (default: 5, or 20 for --files)
--all - Return all matches (use with --min-score to filter)
--min-score <num> - Minimum similarity score
--full - Output full document instead of snippet
--line-numbers - Add line numbers to output
--files - Output docid,score,filepath,context (default: 20 results)
--json - JSON output with snippets (default: 20 results)
--csv - CSV output with snippets
--md - Markdown output
--xml - XML output
-c, --collection <name> - Filter results to a specific collection

Multi-get options:
-l <num> - Maximum lines per file
--max-bytes <num> - Skip files larger than N bytes (default: 10240)
--json/--csv/--md/--xml/--files - Output format (same as search)

Models (auto-downloaded from HuggingFace):
Embedding: embeddinggemma-300M-Q8_0
Reranking: qwen3-reranker-0.6b-q8_0
Generation: Qwen3-0.6B-Q8_0

Index: /home/totaluser/.cache/qmd/index.sqlite
totaluser@Total-Hardware:~$

meager moonBOT Apr 13, 2026, 3:50 AM

#

Ah, yep, this makes the picture a lot clearer.

Your config is fine:

memory.qmd.paths is set correctly
channel scope is set correctly

The failure is lower-level:

qmd --version doesn’t even return a version, just usage
the help output only shows the older --mask collection flow
OpenClaw is trying to manage collections like memory-alt-main, but QMD state is inconsistent, so it falls back to builtin

That’s why you keep seeing default-memory results.

My best read

#

This looks like a QMD compatibility/stale-state problem, not a bad OpenClaw config.

Especially because your log shows:

qmd search ... -c memory-alt-main failed: Collection not found: memory-alt-main
qmd memory failed; switching to builtin index

So OpenClaw is definitely not successfully staying on QMD for that search.

I’d do this in order

1) Update QMD

If you installed it with pnpm, reinstall/upgrade it:

#

pnpm add -g @tobilu/qmd@latest

2) Reset the broken per-agent QMD state, with a backup

For main first:

openclaw gateway stop
mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
openclaw gateway start

3) Rebuild and retest

openclaw memory status --agent main --deep --index --verbose

#

openclaw memory search --agent main --query "quantum mechanics" --json

If that fixes main, repeat for any other affected agents.

One extra note

Plain qmd in your shell is using:

Index: /home/totaluser/.cache/qmd/index.sqlite

But OpenClaw is using the per-agent managed index under:

#

~/.openclaw/agents/main/qmd/...

So the important reset is the OpenClaw agent QMD dir, not the default shell QMD cache.

If this still breaks after:

updating QMD, and
resetting ~/.openclaw/agents/main/qmd

then I’d call it a real bug and file it here with those logs:
https://github.com/openclaw/openclaw/issues

I’m pretty confident we’ve narrowed it down now.

true hound Apr 13, 2026, 4:05 AM

#

ok i updated qmd now what

meager moonBOT Apr 13, 2026, 4:06 AM

#

Nice. Next do the clean per-agent QMD reset, then rebuild:

openclaw gateway stop
mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
openclaw gateway start
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "quantum mechanics" --json

What this does:

keeps a backup
forces OpenClaw to recreate the managed QMD collections
tests whether main now searches the extra path instead of falling back

If main works after that, you can repeat the same pattern for any other affected agents.

#

Paste the last two command outputs if it still acts weird.

true hound Apr 13, 2026, 4:09 AM

#

can i wipe qmd off of my device and then just have openclaw handle all of it?

meager moonBOT Apr 13, 2026, 4:09 AM

#

Not if you want to keep using the QMD backend.

OpenClaw can manage the per-agent QMD state for you, but it still needs the qmd binary installed.

So:

Yes: wipe ~/.openclaw/agents/<agent>/qmd and let OpenClaw recreate it
No: uninstall qmd entirely and still expect memory.backend: "qmd" to work

If you remove the binary, OpenClaw will either fail QMD or fall back to builtin memory.

Best option

Keep QMD installed, but reset only the managed state:

openclaw gateway stop

#

mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
openclaw gateway start

If you want zero QMD dependency

Then switch to builtin memory instead:

{
  memory: {
    backend: "builtin"
  }
}

But then you lose the QMD extra-path behavior.

#

So my recommendation is: keep the qmd binary, wipe the agent qmd state only.

true hound Apr 13, 2026, 4:10 AM

#

can i completely wipe qmd and reinstall it?

meager moonBOT Apr 13, 2026, 4:10 AM

#

Yep, you can. I’d do it the safe way, with backups instead of deleting first.

For the agent you’re testing (main), this is a clean reset:

openclaw gateway stop
pnpm remove -g @tobilu/qmd
mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
mv ~/.cache/qmd ~/.cache/qmd.bak.$(date +%s)
pnpm add -g @tobilu/qmd@latest
openclaw gateway start
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "quantum mechanics" --json

A couple notes:

#

~/.openclaw/agents/main/qmd is the important OpenClaw-managed state
~/.cache/qmd is the plain shell QMD cache, optional but fine to reset too
using mv keeps it recoverable if anything weird happens

If main works after that, you can repeat the same ~/.openclaw/agents/<agent>/qmd reset for other agents.

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it for the next person.

true hound Apr 13, 2026, 4:31 AM

#

is this right

openclaw config set memory.qmd.command "/home/totaluser/.npm-global/bin/qmd"

meager moonBOT Apr 13, 2026, 4:32 AM

#

Yes, memory.qmd.command is the right config key.

But only set it to that path if it’s the actual executable path.

Best way:

command -v qmd

If that prints:

/home/totaluser/.npm-global/bin/qmd

then this is correct:

#

openclaw config set memory.qmd.command "/home/totaluser/.npm-global/bin/qmd"

You can also do it directly from the resolved path:

openclaw config set memory.qmd.command "$(command -v qmd)"

Then restart and retest:

openclaw gateway restart
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "quantum mechanics" --json

#

If command -v qmd prints a different path, use that instead.

true hound Apr 13, 2026, 4:40 AM

#

ok so i indexed and then i did qmd collection list and it says thre are no collections. doesnt that mean its not working??

meager moonBOT Apr 13, 2026, 4:40 AM

#

Not necessarily.

Plain:

qmd collection list

checks your shell’s default QMD home, usually ~/.cache/qmd.

But OpenClaw uses a per-agent QMD home under:

~/.openclaw/agents/main/qmd

So if plain qmd collection list says “no collections”, that can just mean:

#

your default shell QMD index has no collections
while OpenClaw’s managed agent QMD index is a different one

To inspect the same QMD state OpenClaw uses for main, run:

XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache" \
qmd collection list

And this is also useful:

XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config" \

#

QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache" \
qmd status

So:

plain qmd collection list empty → not enough evidence
agent-scoped qmd collection list empty after indexing → yeah, then something is still wrong

Paste the output of the agent-scoped qmd collection list and I can tell you which side it’s failing on.

true hound Apr 13, 2026, 4:53 AM

#

ok so running this

totaluser@Total-Hardware:~$ XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config"
QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd"
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache"
qmd collection list
Collections (4):

memory-dir-main (qmd://memory-dir-main/)
Pattern: **/*.md
Files: 62
Updated: 5h ago

yr-knowledge (qmd://yr-knowledge/)
Pattern: **/*.md
Files: 38
Updated: 6h ago

sessions-main (qmd://sessions-main/)
Pattern: **/*.md
Files: 198
Updated: 3m ago

memory-root-main (qmd://memory-root-main/)
Pattern: **/*.md
Files: 862
Updated: 5h ago

is very hopeful becaise yrknowledge is the extra path im trying to get it to index. so this is good news but when i try to search for something that i know is in the yr-knowledge i get a list of qmd fails i will send it right after this

#

totaluser@Total-Hardware:~$ openclaw memory search --agent main --query "thunderbird"
21:49:06-07:00 [plugins] plugins.allow is empty; discovered non-bundled plugins may auto-load: lossless-claw (/home/totaluser/.openclaw/extensions/lossless-claw/dist/index.js). Set plugins.allow to explicit trusted ids.

🦞 OpenClaw 2026.4.11 (769908e)
Automation with claws: minimal fuss, maximal pinch.

21:49:18-07:00 [memory] qmd collection add skipped for memory-alt-main: qmd collection add /home/totaluser/.openclaw/workspace --name memory-alt-main --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-main (qmd://memory-root-main/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-main'

21:49:19-07:00 [memory] qmd search failed because a managed collection is missing; repairing collections and retrying once
21:49:23-07:00 [memory] qmd collection add skipped for memory-alt-main: qmd collection add /home/totaluser/.openclaw/workspace --name memory-alt-main --glob memory.md failed (code 1): A collection already exists for this path and pattern:
Name: memory-root-main (qmd://memory-root-main/)
Pattern: **/*.md

Use 'qmd update' to re-index it, or remove it first with 'qmd collection remove memory-root-main'

21:49:24-07:00 [memory] qmd search failed: Error: qmd search thunderbird --json -n 6 -c memory-alt-main failed (code 1): Collection not found: memory-alt-main

21:49:24-07:00 [memory] qmd memory failed; switching to builtin index: qmd search thunderbird --json -n 6 -c memory-alt-main failed (code 1): Collection not found: memory-alt-main

No matches.

meager moonBOT Apr 13, 2026, 4:54 AM

#

Yep, this is enough to call it:

yr-knowledge is indexed
but OpenClaw is still trying to search a missing collection: memory-alt-main
that failure makes it fall back before QMD results can be used

So the extra path itself is not the problem anymore. The problem is the auto-managed default-memory collections.

What I think is happening

Your live QMD collections are:

memory-dir-main
memory-root-main
sessions-main
yr-knowledge

#

But OpenClaw expects an additional one:

memory-alt-main

Since that one does not exist, every QMD search dies here:

Collection not found: memory-alt-main

and then builtin fallback takes over.

Best workaround right now

Disable the automatic default-memory collection setup and define the paths explicitly.

Something like this in config:

#

{
  memory: {
    backend: "qmd",
    qmd: {
      includeDefaultMemory: false,
      paths: [
        { name: "workspace-memory-root", path: "~/.openclaw/workspace", pattern: "MEMORY.md" },
        { name: "workspace-memory-dir", path: "~/.openclaw/workspace/memory", pattern: "**/*.md" },
        { name: "yr-knowledge", path: "~/.openclaw/yr-knowledge", pattern: "**/*.md" }
      ],
      sessions: { enabled: true }
    }
  }
}

#

Then rebuild the agent QMD state again:

openclaw gateway stop
mv ~/.openclaw/agents/main/qmd ~/.openclaw/agents/main/qmd.bak.$(date +%s)
openclaw gateway start
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "thunderbird"

Short version

Good news: your extra path works.
Bad news: OpenClaw’s default QMD collection management is the broken part.

At this point I’d also file a bug with these logs:
https://github.com/openclaw/openclaw/issues

#

If you want, I can help you turn that config into exact openclaw config set ... commands next.

true hound Apr 13, 2026, 4:56 AM

#

no but i have multiple agents can i jsut force create the alt collection so that it is there?

meager moonBOT Apr 13, 2026, 4:56 AM

#

I wouldn’t go that route.

Two reasons:

I don’t see a real force-create flag in the QMD CLI you pasted.
Manual qmd collection add ... --name memory-alt-main --mask memory.md will probably just hit the same conflict again, because memory-root-main is already overlapping it.

So, practically: no, not in a clean/reliable way.

With multiple agents, manual collection surgery is going to be brittle and annoying. OpenClaw may just run back into the same managed-collection mismatch on the next sync.

Better fix for multi-agent setups

Use the workaround once in config:

memory.qmd.includeDefaultMemory: false
define the memory roots you actually want under memory.qmd.paths

#

That avoids the broken auto-created memory-alt-* path entirely.

Then reset the per-agent QMD state for all agents.

If you want, I can give you:

the exact config block for your setup, and
one safe shell loop to back up every ~/.openclaw/agents/*/qmd dir at once.

true hound Apr 13, 2026, 4:57 AM

#

but right now each agent has their own memory qmd database right?

meager moonBOT Apr 13, 2026, 4:58 AM

#

Yes.

With QMD, OpenClaw gives each agent its own QMD state and index DB under:

~/.openclaw/agents/<agentId>/qmd/

So for example:

~/.openclaw/agents/main/qmd/...
~/.openclaw/agents/forge/qmd/...
~/.openclaw/agents/scout/qmd/...

That means:

each agent has its own collections

#

each agent has its own index.sqlite
each agent can index the same external path, but it still ends up in that agent’s own QMD DB

So yes, right now they’re separate, per-agent databases.

true hound Apr 13, 2026, 4:59 AM

#

and if i dont include defualt memory wont that break that feature where each agent can have their own memry collection

meager moonBOT Apr 13, 2026, 4:59 AM

#

No, not if you add the workspace memory paths back explicitly.

includeDefaultMemory: false only means:

“don’t auto-create the built-in default collections for me”

It does not mean:

“all agents lose separate memory”

Per-agent separation still exists because each agent still has its own workspace and own QMD DB.

The important part is this:

if you set includeDefaultMemory: false and add nothing back, then yes, you’d stop indexing the normal per-agent workspace memory files
if you set includeDefaultMemory: false but explicitly add MEMORY.md and memory/**/*.md, then each agent still keeps its own memory, just without the broken auto-created memory-alt-* collection

#

A good multi-agent-friendly version is:

{
  memory: {
    backend: "qmd",
    qmd: {
      includeDefaultMemory: false,
      paths: [
        { name: "workspace-memory-root", path: ".", pattern: "MEMORY.md" },
        { name: "workspace-memory-dir", path: "memory", pattern: "**/*.md" },
        { name: "yr-knowledge", path: "/home/totaluser/.openclaw/yr-knowledge", pattern: "**/*.md" }
      ],
      sessions: { enabled: true }
    }
  }

#

Why this works:

. and memory resolve relative to each agent’s own workspace
so main, forge, scout, etc. still each get their own memory collections
yr-knowledge is the shared extra path layered on top

So: agent-specific memory stays, you just bypass the broken default auto-collection logic.

true hound Apr 13, 2026, 5:01 AM

#

ok lets try it

meager moonBOT Apr 13, 2026, 5:01 AM

#

Cool, let’s do the workaround cleanly.

Run these:

openclaw config set memory.qmd.includeDefaultMemory false --strict-json
openclaw config set memory.qmd.paths '[{"name":"workspace-memory-root","path":".","pattern":"MEMORY.md"},{"name":"workspace-memory-dir","path":"memory","pattern":"**/*.md"},{"name":"yr-knowledge","path":"/home/totaluser/.openclaw/yr-knowledge","pattern":"**/*.md"}]' --strict-json
openclaw gateway stop
for d in "$HOME"/.openclaw/agents/*/qmd; do [ -d "$d" ] && mv "$d" "$d.bak.$(date +%s)"; done
openclaw gateway start
openclaw memory status --agent main --deep --index --verbose
openclaw memory search --agent main --query "thunderbird"

What this does:

disables the broken auto-default QMD collections

#

adds back each agent’s own MEMORY.md and memory/**/*.md
keeps your shared yr-knowledge path
resets every agent’s QMD state so collections get recreated from the new config

Paste the last two outputs and we’ll check if it finally behaves.

true hound Apr 13, 2026, 5:09 AM

#

ok so it worked! it actually is searching the extra path!

meager moonBOT Apr 13, 2026, 5:10 AM

#

Hell yes, glad we caught it 🎉

So the extra path was fine, and the real problem was the auto default QMD collection setup tripping over memory-alt-*.

I’d keep that workaround config for now.

If you’re up for it, this looks worth filing as a bug:
https://github.com/openclaw/openclaw/issues

Useful summary for the issue:

memory.qmd.paths extra path was indexed correctly
OpenClaw kept expecting missing memory-alt-main
QMD search failed and fell back to builtin
workaround was memory.qmd.includeDefaultMemory: false plus explicit paths for:
- . with MEMORY.md
- memory with **/*.md
- shared extra path

#

And if this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it for the next person.

true hound Apr 13, 2026, 5:12 AM

#

how do i force a reindex for main

meager moonBOT Apr 13, 2026, 5:13 AM

#

Use:

openclaw memory index --agent main --force --verbose

That forces a full reindex for main.

If you want to verify after:

openclaw memory status --agent main --deep

Glad it’s finally behaving 🦐

true hound Apr 13, 2026, 5:20 AM

#

after a search through cli using qmd it freezes after pulling up the results

meager moonBOT Apr 13, 2026, 5:21 AM

#

If you mean the plain qmd CLI, that sounds more like a QMD-side terminal/pager issue than an OpenClaw one.

Quick checks:

Try q
If it drops you back to the prompt, it was probably sitting in a pager.
Try non-interactive output:

qmd search "thunderbird" --json

or even:

qmd search "thunderbird" --json | cat

#

If you want the OpenClaw path specifically, use:

openclaw memory search --agent main --query "thunderbird" --json

If it still “freezes”, paste the exact command you ran and whether:

q exits
Ctrl+C exits
it happens with --json too

If it’s only qmd ... and not openclaw memory search ..., I’d treat that as a QMD CLI issue, not the OpenClaw memory config issue.

true hound Apr 13, 2026, 5:45 AM

#

ok it works!! i want to move yr-knowledge into .openclaw/rush/

i also want to add these folders to be added to the global qmd.

.openclaw/workspace/health_research and .openclaw/workspace/research

meager moonBOT Apr 13, 2026, 5:46 AM

#

Nice 😄

You can do that, but I’d honestly recommend keeping shared knowledge outside rush/ if it’s meant to be global. Putting it under one agent’s workspace works, but it’s a little confusing later.

If you still want it there, update the global QMD paths to absolute paths like this:

openclaw config set memory.qmd.paths '[{"name":"workspace-memory-root","path":".","pattern":"MEMORY.md"},{"name":"workspace-memory-dir","path":"memory","pattern":"**/*.md"},{"name":"yr-knowledge","path":"/home/totaluser/.openclaw/rush/yr-knowledge","pattern":"**/*.md"},{"name":"health-research","path":"/home/totaluser/.openclaw/workspace/health_research","pattern":"**/*.md"},{"name":"research","path":"/home/totaluser/.openclaw/workspace/research","pattern":"**/*.md"}]' --strict-json

Then reindex all agents:

openclaw memory index --force --verbose

Because without --agent, it runs across every configured agent.

#

My recommendation

If it’s truly shared/global, use:

~/.openclaw/yr-knowledge
~/.openclaw/workspace/health_research
~/.openclaw/workspace/research

If you want, I can give you the recommended version of that command with yr-knowledge left in the shared location instead of moving it into rush/.

true hound Apr 13, 2026, 5:47 AM

#

does adding a ton more files into qmd searches affect retrieval speed? and when it comes to health research, i really only need one agent maybe two to be searching through that how can i do that with our current set up and also is this something that makes sense to go under memory qmd or should it be a separate qmd collection?

meager moonBOT Apr 13, 2026, 5:48 AM

#

Yes, adding a lot more files can affect things:

indexing time goes up
disk usage goes up
search latency can go up too, especially if you make every agent index the same large shared corpus

And in your setup, that cost multiplies, because each agent has its own QMD DB.

My recommendation

For stuff like health_research, if only one or two agents need it, do not put it in global memory.qmd.paths.

Instead, add it only to those agents with:

agents.list[].memorySearch.qmd.extraCollections

That gives you a separate QMD collection for just those agents, inside their own per-agent DBs.

#

When to use which

Use `memory.qmd.paths`

For docs that are truly shared across most/all agents.

Use `agents.list[].memorySearch.qmd.extraCollections`

For docs that only specific agents should search, like health research.

Does separate collection make sense?

Yes, definitely.

For health_research, a separate collection is the right shape.
You do not need a whole separate QMD install or separate global DB, just a separate collection attached to the relevant agents.

Practical rule

Shared by everybody → memory.qmd.paths
Only for 1-2 agents → agents.list[].memorySearch.qmd.extraCollections

#

So for your case, I’d keep:

global/shared stuff in memory.qmd.paths
health_research as a per-agent extra collection for the health-focused agent(s)

If you want, I can write the exact config snippet for, say, vital and one second agent next.

true hound Apr 13, 2026, 5:53 AM

#

ok so in that case i really only need rush to have access to yr knowledge, and main, scout, and vital to have the health research one.

research can be accessed by everyone.

where should these folders be at in the file structure? yr knowledge has to be in rushs workspace

meager moonBOT Apr 13, 2026, 5:54 AM

#

I’d organize it like this:

~/.openclaw/
  shared/
    health_research/
    research/
  rush/
    yr-knowledge/

Why this layout

agent-local stuff lives in that agent’s workspace
shared stuff lives in a neutral shared folder, not inside one agent’s workspace

So for your case:

#

yr-knowledge → ~/.openclaw/rush/yr-knowledge/ ✅
health_research → ~/.openclaw/shared/health_research/
research → ~/.openclaw/shared/research/

How I’d wire it

Global, for everyone

Put research in global memory.qmd.paths

Rush only

Put yr-knowledge in rush’s memorySearch.qmd.extraCollections

Since it’s in Rush’s workspace, you can even reference it as:

{ path: "yr-knowledge" }

#

Main + Scout + Vital only

Put health_research in those three agents’ memorySearch.qmd.extraCollections, using the same absolute shared path:

{ path: "/home/totaluser/.openclaw/shared/health_research", name: "health-research" }

Rule of thumb

inside an agent workspace = that agent’s stuff
outside agent workspaces, in a shared dir = shared corpus for selected agents

So yes, if yr-knowledge has to be in Rush’s workspace, that makes sense.
But I would not put health_research or research inside workspace/ if multiple agents need them. A neutral shared/ folder is cleaner.

If you want, I can write the exact config snippet for:

global research
Rush-only yr-knowledge

#

Main/Scout/Vital-only health_research

true hound Apr 13, 2026, 5:55 AM

#

yeah ill handle putting the folders in the right spot can you give me the config snippet for this

meager moonBOT Apr 13, 2026, 5:55 AM

#

Yep, here’s a clean config shape for that setup:

{
  memory: {
    backend: "qmd",
    qmd: {
      includeDefaultMemory: false,
      paths: [
        { name: "workspace-memory-root", path: ".", pattern: "MEMORY.md" },
        { name: "workspace-memory-dir", path: "memory", pattern: "**/*.md" },

        // shared by everyone
        { name: "research", path: "/home/totaluser/.openclaw/shared/research", pattern: "**/*.md" },
      ],
      sessions: { enabled: true },

#

      scope: {
        default: "deny",
        rules: [
          { action: "allow", match: { chatType: "direct" } },
          { action: "allow", match: { chatType: "channel" } },
        ],
      },
    },
  },

  agents: {
    list: [
      {
        id: "rush",
        memorySearch: {

#

          qmd: {
            extraCollections: [
              // relative to rush workspace
              { path: "yr-knowledge" },
            ],
          },
        },
      },

      {
        id: "main",
        memorySearch: {
          qmd: {
            extraCollections: [
              { path: "/home/totaluser/.openclaw/shared/health_research", name: "health-research" },

#

            ],
          },
        },
      },

      {
        id: "scout",
        memorySearch: {
          qmd: {
            extraCollections: [
              { path: "/home/totaluser/.openclaw/shared/health_research", name: "health-research" },
            ],
          },
        },
      },

#


      {
        id: "vital",
        memorySearch: {
          qmd: {
            extraCollections: [
              { path: "/home/totaluser/.openclaw/shared/health_research", name: "health-research" },
            ],
          },
        },
      },
    ],
  },
}

#

Two important notes:

If those agents already exist in agents.list, merge this into their existing entries, don’t create duplicate ids.
After updating config, rebuild:

openclaw memory index --force --verbose

If you want, I can also turn this into exact openclaw config set ... commands, but honestly for this amount of nested config, editing the config file is cleaner.

true hound Apr 13, 2026, 5:56 AM

#

can you give me the config set commands

meager moonBOT Apr 13, 2026, 5:56 AM

#

Yep. First, set the global stuff:

openclaw config set memory.qmd.includeDefaultMemory false --strict-json

openclaw config set memory.qmd.paths '[{"name":"workspace-memory-root","path":".","pattern":"MEMORY.md"},{"name":"workspace-memory-dir","path":"memory","pattern":"**/*.md"},{"name":"research","path":"/home/totaluser/.openclaw/shared/research","pattern":"**/*.md"}]' --strict-json

Then set the per-agent collections.

You’ll need the array indexes for each agent in agents.list, so first check:

openclaw config get agents.list --json

Then use the right indexes in these commands:

#

# Rush only
openclaw config set agents.list[RUSH_INDEX].memorySearch.qmd.extraCollections '[{"path":"yr-knowledge"}]' --strict-json

# Main only
openclaw config set agents.list[MAIN_INDEX].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

# Scout only
openclaw config set agents.list[SCOUT_INDEX].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

# Vital only
openclaw config set agents.list[VITAL_INDEX].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

Then rebuild:

#

openclaw memory index --force --verbose

Two important gotchas:

these extraCollections commands replace that agent’s current extraCollections array
if any of those agents already have other extra collections you want to keep, include them in the JSON too

If you paste the output of:

openclaw config get agents.list --json

I can turn those RUSH_INDEX / MAIN_INDEX / etc into exact commands.

true hound Apr 13, 2026, 5:59 AM

#

can you give me the code with the index for these

totaluser@Total-Hardware:~$ openclaw config get agents.list --json
[
{
"id": "main",
"heartbeat": {},
"subagents": {
"allowAgents": [
"forge",
"scout",
"vital",
"sentinel",
"alfred"
]
}
},
{
"id": "forge",
"name": "Forge",
"workspace": "/home/totaluser/.openclaw/forge",
"agentDir": "/home/totaluser/.openclaw/agents/forge/agent",
"heartbeat": {}
},
{
"id": "scout",
"name": "Scout",
"workspace": "/home/totaluser/.openclaw/scout",
"agentDir": "/home/totaluser/.openclaw/agents/scout/agent",
"heartbeat": {}
},
{
"id": "vital",
"name": "vital",
"workspace": "/home/totaluser/.openclaw/vital",
"agentDir": "/home/totaluser/.openclaw/agents/vital/agent",
"heartbeat": {}
},
{
"id": "sentinel",
"name": "sentinel",
"workspace": "/home/totaluser/.openclaw/sentinel",
"agentDir": "/home/totaluser/.openclaw/agents/sentinel/agent",
"heartbeat": {}
},
{
"id": "alfred",
"name": "Alfred",
"workspace": "/home/totaluser/.openclaw/alfred",
"agentDir": "/home/totaluser/.openclaw/agents/alfred/agent",
"heartbeat": {
"every": "15m",
"target": "telegram",
"accountId": "forge"
}
},
{
"id": "rush",
"name": "rush",
"workspace": "/home/totaluser/.openclaw/rush",
"agentDir": "/home/totaluser/.openclaw/agents/rush/agent",
"memorySearch": {
"qmd": {
"extraCollections": []
}
}
},
{
"id": "cipher",
"name": "Cipher",
"workspace": "/home/totaluser/.openclaw/cipher",
"agentDir": "/home/totaluser/.openclaw/agents/cipher/agent"
}
]
totaluser@Total-Hardware:~$

meager moonBOT Apr 13, 2026, 5:59 AM

#

Yep, here are the exact ones for your indexes:

main = 0
scout = 2
vital = 3
rush = 6

Run these:

openclaw config set memory.qmd.includeDefaultMemory false --strict-json

openclaw config set memory.qmd.paths '[{"name":"workspace-memory-root","path":".","pattern":"MEMORY.md"},{"name":"workspace-memory-dir","path":"memory","pattern":"**/*.md"},{"name":"research","path":"/home/totaluser/.openclaw/shared/research","pattern":"**/*.md"}]' --strict-json

openclaw config set agents.list[6].memorySearch.qmd.extraCollections '[{"path":"yr-knowledge"}]' --strict-json

#

openclaw config set agents.list[0].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

openclaw config set agents.list[2].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

openclaw config set agents.list[3].memorySearch.qmd.extraCollections '[{"path":"/home/totaluser/.openclaw/shared/health_research","name":"health-research"}]' --strict-json

openclaw memory index --force --verbose

That should give you:

research for everyone
yr-knowledge for rush only
health_research for main, scout, and vital

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

true hound Apr 13, 2026, 6:06 AM

#

how often does openclaw reindex

meager moonBOT Apr 13, 2026, 6:06 AM

#

For QMD, by default OpenClaw reindexes:

on boot
then every 5 minutes

It also debounces file changes by about 15 seconds.

The knobs are:

memory.qmd.update.interval → default 5m
memory.qmd.update.debounceMs → default 15000
memory.qmd.update.onBoot → default true

So in practice, with your setup, think:

startup + every 5 minutes + short debounce after changes

#

And if you want to force it manually:

openclaw memory index --agent main --force --verbose

Glad that helped! 🎉
If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution.
This helps Answer Overflow index the solution for future users.

ripe ruinBOT Apr 13, 2026, 6:14 AM

#

true hound Apr 13, 2026, 6:35 AM

#

how can i make sure active memory is working

meager moonBOT Apr 13, 2026, 6:36 AM

#

Best way to verify it:

In the chat/session you care about, run:

/active-memory status

Turn on verbose for that session:

/verbose on

Send a normal message, then look for lines like:

🧩 Active Memory: ok ...
🔎 Active Memory Debug: ...

#

That’s the clearest sign it actually ran.

If it still doesn’t show up, check these

plugins.entries.active-memory.enabled is on
the current agent id is listed in plugins.entries.active-memory.config.agents
you’re testing in an interactive persistent chat session
memory search itself is healthy:

openclaw memory status --deep

Good sanity-check flow

#

/active-memory status
/verbose on
what should i remember about my health research setup?

If active memory is working, you should see the normal reply plus the active-memory status/debug footer.

If you want, I can help you check the exact config keys for your setup next.

true hound Apr 13, 2026, 6:39 AM

#

so i asked it this
hey what are some ai facts that we have researched?

OpenClaw-specific:

• v2026.4.x had significant breaking changes: nodes.run shell wrapper removed, dangerous-code fail-closed shipped, background tasks now SQLite-backed
• April 2026 had a rough security stretch — CVE-2026-33579 (privilege escalation), CVE-2026-33581 (media LFI), axios supply chain attack
• China adopted OpenClaw fast — government/explore limits, Ant Group/Tsinghua partnered on security paper, ClawHub launched

AI Agents:

• Context engineering is replacing RAG at scale — longer context windows eliminate the need for retrieval
• MCP (Model Context Protocol) became the standard bridge for agent-to-tool communication
• OpenClaw has ~15+ agent frameworks competing — LangChain, CrewAI, AutoGen, etc.
• Claude Code launched and became the main OpenClaw competitor in the AI coding space
• Evoclaw benchmark — tests agent coherence across long tasks, agents still struggle past ~20 tool calls

AI Landscape:

• Agentic AI went from "hype" to "real deployment" in 6 months (March 2026)

it looks like its not using qmd

meager moonBOT Apr 13, 2026, 6:40 AM

#

Yeah, that specific turn did not use QMD. It used exec.

But that means tool choice, not necessarily QMD failure.

What’s going on

There are 3 different things here:

QMD / memory search works
- proven by openclaw memory search ... working earlier
Active Memory
- a hidden pre-pass that can inject memory context
- does not guarantee the agent won’t also use other tools
Exec
- if the agent has it, it may decide raw file listing/grep is easier than memory search

#

So the run you pasted says:

“the agent chose exec for this question”

not

“QMD is broken”

How to tell if Active Memory is actually running

In that session:

/active-memory status
/verbose on

#

Then ask another question and look for:

🧩 Active Memory: ok ...
🔎 Active Memory Debug: ...

If you see that, Active Memory ran.

If you want it to prefer QMD more often

Add an instruction like this to the agent’s AGENTS.md or persona prompt:

For questions about research, notes, or prior findings, use memory search / memory get before exec or raw filesystem scans unless memory search fails.

Because right now, with exec available, the model is just taking the brute-force path.

#

Short version

Yes, that reply used exec, not QMD
No, that does not prove QMD is broken
it proves the agent preferred exec for that question

If you want, I can help you write a short instruction so the agent uses memory search first for research questions.

true hound Apr 13, 2026, 6:42 AM

#

give me that short thing and then once specific for rush saying that for any context or info that rush needs about the program to use memory search. and any time i ask for info or to be reminded of something he has to use memory

meager moonBOT Apr 13, 2026, 6:42 AM

#

Yep, here are two short versions you can drop in.

Generic

For questions about research, notes, prior findings, stored context, or reminders, use memory search / memory get before exec or raw filesystem scans. Only fall back to exec if memory search fails or does not return enough context.

Rush-specific

For Rush: whenever Peter asks for context, prior program knowledge, reminders, what we know, what we researched, or anything Rush should remember about the program, use memory search / memory get first. Do not answer from guesswork or raw file scanning unless memory search fails. If Peter asks to be reminded of something or asks what we know about a topic, Rush must use memory tools.

If you want, I can make that even tighter and more forceful for AGENTS.md, like a “memory-first policy” version.

true hound Apr 13, 2026, 6:43 AM

#

ok but how do i make sure active memory is on and being used

meager moonBOT Apr 13, 2026, 6:44 AM

#

To make sure Active Memory is actually on and can run in this Discord thread, I’d check all 4 gates:

1) Plugin enabled

openclaw config set plugins.entries.active-memory.enabled true --strict-json

2) Global Active Memory toggle enabled

openclaw config set plugins.entries.active-memory.config.enabled true --strict-json

3) Current agent targeted

Example for rush:

#

openclaw config set plugins.entries.active-memory.config.agents '["rush"]' --strict-json

If you want multiple agents, include them all:

openclaw config set plugins.entries.active-memory.config.agents '["main","scout","vital","rush"]' --strict-json

4) Allowed chat type includes channel

This part is important because you’re testing in a Discord thread/channel, and the default is usually direct-only.

openclaw config set plugins.entries.active-memory.config.allowedChatTypes '["direct","channel"]' --strict-json

Optional but very helpful while testing

#

openclaw config set plugins.entries.active-memory.config.logging true --strict-json
openclaw gateway restart

Then in the chat session itself:

/active-memory on
/active-memory status
/verbose on

Then ask a question and look for lines like:

#

🧩 Active Memory: ok ...
🔎 Active Memory Debug: ...

If you see those, it’s being used.

Short version

For your case, the biggest gotcha is probably:

Active Memory won’t run in this thread unless allowedChatTypes includes "channel"

If you want, I can also give you one compact config snippet just for Rush in channels.

true hound Apr 13, 2026, 6:45 AM

#

yeah can you give me the code snippet for you to check what my active memory config is

meager moonBOT Apr 13, 2026, 6:45 AM

#

Yep, run this and paste the output:

openclaw config get plugins.entries.active-memory --json

And if you want the full picture, these are useful too:

openclaw config get plugins.entries.active-memory --json
openclaw config get plugins.slots.memory --json
openclaw config get agents.list --json

If you want to check it from inside the chat session too:

#

/active-memory status
/active-memory status --global

Paste the plugins.entries.active-memory output and I’ll tell you exactly what’s on/off.

true hound Apr 13, 2026, 6:46 AM

#

totaluser@Total-Hardware:~$ openclaw config get plugins.entries.active-memory --json
{
"enabled": true,
"config": {
"agents": [
"main",
"forge",
"rush",
"scout",
"vital"
],
"allowedChatTypes": [
"direct",
"channel"
],
"queryMode": "recent",
"promptStyle": "balanced",
"timeoutMs": 15000,
"maxSummaryChars": 220,
"logging": true
}
}

here is part 1 of 2

#

totaluser@Total-Hardware:~$ openclaw config get plugins.entries.active-memory --json
openclaw config get plugins.slots.memory --json
openclaw config get agents.list --json
{
"enabled": true,
"config": {
"agents": [
"main",
"forge",
"rush",
"scout",
"vital"
],
"allowedChatTypes": [
"direct",
"channel"
],
"queryMode": "recent",
"promptStyle": "balanced",
"timeoutMs": 15000,
"maxSummaryChars": 220,
"logging": true
}
}
Config path not found: plugins.slots.memory
[
{
"id": "main",
"memorySearch": {
"qmd": {
"extraCollections": [
{
"path": "/home/totaluser/.openclaw/shared/health_research",
"name": "health-research"
}
]
}
},
"heartbeat": {},
"subagents": {
"allowAgents": [
"forge",
"scout",
"vital",
"sentinel",
"alfred"
]
}
},
{
"id": "forge",
"name": "Forge",
"workspace": "/home/totaluser/.openclaw/forge",
"agentDir": "/home/totaluser/.openclaw/agents/forge/agent",
"heartbeat": {}
},
{
"id": "scout",
"name": "Scout",
"workspace": "/home/totaluser/.openclaw/scout",
"agentDir": "/home/totaluser/.openclaw/agents/scout/agent",
"memorySearch": {
"qmd": {
"extraCollections": [
{
"path": "/home/totaluser/.openclaw/shared/health_research",
"name": "health-research"
}
]
}
},
"heartbeat": {}
},
{
"id": "vital",
"name": "vital",
"workspace": "/home/totaluser/.openclaw/vital",
"agentDir": "/home/totaluser/.openclaw/agents/vital/agent",
"memorySearch": {
"qmd": {
"extraCollections": [
{
"path": "/home/totaluser/.openclaw/shared/health_research",
"name": "health-research"
}
]
}
},
"heartbeat": {}
},
sorry this is part 2/3

#

{
"id": "sentinel",
"name": "sentinel",
"workspace": "/home/totaluser/.openclaw/sentinel",
"agentDir": "/home/totaluser/.openclaw/agents/sentinel/agent",
"heartbeat": {}
},
{
"id": "alfred",
"name": "Alfred",
"workspace": "/home/totaluser/.openclaw/alfred",
"agentDir": "/home/totaluser/.openclaw/agents/alfred/agent",
"heartbeat": {
"every": "15m",
"target": "telegram",
"accountId": "forge"
}
},
{
"id": "rush",
"name": "rush",
"workspace": "/home/totaluser/.openclaw/rush",
"agentDir": "/home/totaluser/.openclaw/agents/rush/agent",
"memorySearch": {
"qmd": {
"extraCollections": [
{
"path": "yr-knowledge"
}
]
}
}
},
{
"id": "cipher",
"name": "Cipher",
"workspace": "/home/totaluser/.openclaw/cipher",
"agentDir": "/home/totaluser/.openclaw/agents/cipher/agent"
}
]
totaluser@Total-Hardware:~$

part 3/3

meager moonBOT Apr 13, 2026, 6:47 AM

#

This actually looks pretty close to correct.

What looks good

plugin is enabled: enabled: true
targeted agents include: main, forge, rush, scout, vital
allowed chat types include: direct and channel
logging is on

So for those agents, in a Discord thread/channel, Active Memory should be eligible.

One thing I’d still set explicitly

Your config does not show:

config.enabled: true

#

It may still work depending on defaults, but if you want zero ambiguity, set it:

openclaw config set plugins.entries.active-memory.config.enabled true --strict-json
openclaw gateway restart

Also, this is not a problem

Config path not found: plugins.slots.memory

That’s fine. It does not mean Active Memory is broken.

So my read is

For these agents:

#

main
forge
rush
scout
vital

Active Memory config looks basically right, as long as you test in one of those agents.

It will not run for:

sentinel
alfred
cipher

because they are not in config.agents.

Best test now

#

In the actual session you care about:

/active-memory on
/active-memory status
/verbose on

Then ask something memory-relevant and watch for:

🧩 Active Memory: ok ...
🔎 Active Memory Debug: ...

If you want the cleanest next step, I’d do exactly this:

#

openclaw config set plugins.entries.active-memory.config.enabled true --strict-json
openclaw gateway restart

Then test again in rush or main.

true hound Apr 14, 2026, 4:24 AM

#

how can i check qmd collections that openclaw uses again?

meager moonBOT Apr 14, 2026, 4:24 AM

#

Use the agent-scoped QMD env, not plain qmd collection list.

For main:

XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache" \
qmd collection list

Template for any agent:

XDG_CONFIG_HOME="$HOME/.openclaw/agents/<agent-id>/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/<agent-id>/qmd/xdg-config/qmd" \

#

XDG_CACHE_HOME="$HOME/.openclaw/agents/<agent-id>/qmd/xdg-cache" \
qmd collection list

And if you also want the matching QMD index status:

XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache" \
qmd status

Plain qmd collection list checks your default shell QMD cache, not OpenClaw’s per-agent one.

true hound Apr 14, 2026, 4:25 AM

#

can you give me rush, main, scout, vital, and forge

meager moonBOT Apr 14, 2026, 4:25 AM

#

Yep, here you go.

rush

XDG_CONFIG_HOME="$HOME/.openclaw/agents/rush/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/rush/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/rush/qmd/xdg-cache" \
qmd collection list

main

XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/main/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache" \
qmd collection list

#

scout

XDG_CONFIG_HOME="$HOME/.openclaw/agents/scout/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/scout/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/scout/qmd/xdg-cache" \
qmd collection list

vital

XDG_CONFIG_HOME="$HOME/.openclaw/agents/vital/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/vital/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/vital/qmd/xdg-cache" \
qmd collection list

forge

#

XDG_CONFIG_HOME="$HOME/.openclaw/agents/forge/qmd/xdg-config" \
QMD_CONFIG_DIR="$HOME/.openclaw/agents/forge/qmd/xdg-config/qmd" \
XDG_CACHE_HOME="$HOME/.openclaw/agents/forge/qmd/xdg-cache" \
qmd collection list

If you want, I can also give you the same 5 as qmd status commands too.

true hound Apr 14, 2026, 4:27 AM

#

what is considered "too large" of a qmd where speed suffers significantly

meager moonBOT Apr 14, 2026, 4:28 AM

#

There isn’t a clean universal cutoff like “past X files QMD is bad”.

In practice, “too large” is when you start seeing one of these:

searches regularly take multiple seconds
you hit QMD timeouts (memory.qmd.limits.timeoutMs, default 4000)
indexing/re-embedding gets annoyingly slow
you duplicate the same big corpus across several agents, so total work/storage multiplies

Practical rule of thumb

Your numbers from earlier, roughly hundreds to low-thousands of chunks per agent, are not huge.

I’d start getting more careful when a single agent’s QMD grows into:

many thousands of chunks, especially on weaker hardware
several large shared corpora indexed into multiple agents

#

DB sizes that keep growing while searches stop feeling “instant enough”

The real bottleneck is usually not raw file count

It’s more about:

total chunk count
number of collections
whether the corpus is duplicated across agents
your machine speed
whether reranking/query expansion is active and how long it takes

Better heuristic

Instead of asking “how big is too big?”, ask:

“Are searches still reliably fast enough for my workflow?”

If yes, you’re fine.

#

If not, the first thing I’d do is:

keep shared corpora only on agents that need them
split specialized corpora into targeted collections
increase timeout if hardware is slower:

openclaw config set memory.qmd.limits.timeoutMs 120000 --strict-json

So short version: your current setup does not sound obviously too large yet.
The bigger danger is indexing too many shared folders into too many agents.

true hound Apr 14, 2026, 4:32 AM

#

how do i check how many chunks a qmd collection for main has

meager moonBOT Apr 14, 2026, 4:32 AM

#

Yep, for main you can query the QMD SQLite DB directly.

This will show chunk/doc counts per collection:

sqlite3 "$HOME/.openclaw/agents/main/qmd/xdg-cache/qmd/index.sqlite" \
'SELECT collection, COUNT(*) AS chunks FROM documents WHERE active = 1 GROUP BY collection ORDER BY chunks DESC;'

That’s the closest “per collection chunk count” check.

If you want it prettier:

sqlite3 -header -column "$HOME/.openclaw/agents/main/qmd/xdg-cache/qmd/index.sqlite" \
'SELECT collection, COUNT(*) AS chunks FROM documents WHERE active = 1 GROUP BY collection ORDER BY chunks DESC;'

#

A quick note:

qmd collection list shows files
this SQLite query shows document/chunk rows per collection

If you want, I can also give you the same command for rush, scout, vital, and forge.

true hound Apr 14, 2026, 4:33 AM

#

so this isnt that much

totaluser@Total-Hardware:~$ sqlite3 -header -column "$HOME/.openclaw/agents/main/qmd/xdg-cache/qmd/index.sqlite"
'SELECT collection, COUNT(*) AS chunks FROM documents WHERE active = 1 GROUP BY collection ORDER BY chunks DESC;'
collection chunks

workspace-memory-root-main 579
research 305
sessions-main 158
workspace-memory-dir-main 63
health-research 62
totaluser@Total-Hardware:~$

meager moonBOT Apr 14, 2026, 4:33 AM

#

Yeah, that’s pretty modest.

For main:

579 chunks in workspace root
305 in research
158 in sessions
63 in memory dir
62 in health research

That is not a scary-sized QMD.

I would not expect that collection mix to cause major speed problems on its own.

If you’re seeing sluggishness, it’s more likely to be from:

the agent choosing a worse tool path

#

QMD query/rerank overhead
disk / machine speed
duplicated corpora across several agents

But purely from chunk count, main looks pretty reasonable.

Rough gut check:

hundreds of chunks, fine
low thousands, usually still fine
many thousands per agent, start watching more closely
same big corpus duplicated across lots of agents, that’s where it gets annoying fast

So yeah, I’d call your current main footprint small to medium, not large.

true hound Apr 14, 2026, 4:34 AM

#

how do i force a memory search through cli again?

meager moonBOT Apr 14, 2026, 4:34 AM

#

Use:

openclaw memory search --agent main --query "your query here" --json

Example:

openclaw memory search --agent main --query "thunderbird" --json

That forces a memory-search run through the CLI for main.

Useful variants:

#

openclaw memory search --agent rush --query "weather research" --json
openclaw memory search --agent scout --query "health facts" --max-results 10 --json

If you want, I can give you a tiny cheat sheet for status, index, and search together.

#cant get extra path for qmd memory to work

What I’d do next

My best read

I’d do this in order

1) Update QMD

2) Reset the broken per-agent QMD state, with a backup

3) Rebuild and retest

One extra note

Best option

If you want zero QMD dependency

What I think is happening

Best workaround right now

Short version

Better fix for multi-agent setups

My recommendation

My recommendation

When to use which

Use memory.qmd.paths

Use agents.list[].memorySearch.qmd.extraCollections

Does separate collection make sense?

Practical rule

Why this layout

How I’d wire it

Global, for everyone

Rush only

Main + Scout + Vital only

Rule of thumb

If it still doesn’t show up, check these

Good sanity-check flow

What’s going on

How to tell if Active Memory is actually running

If you want it to prefer QMD more often

Short version

Generic

Rush-specific

1) Plugin enabled

2) Global Active Memory toggle enabled

3) Current agent targeted

4) Allowed chat type includes channel

Optional but very helpful while testing

Short version

What looks good

One thing I’d still set explicitly

Also, this is not a problem

So my read is

Best test now

rush

main

scout

vital

forge

Practical rule of thumb

The real bottleneck is usually not raw file count

Better heuristic

Use `memory.qmd.paths`

Use `agents.list[].memorySearch.qmd.extraCollections`