Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

After upgrading from 2026.5.26 to 2026.6.1, memorySearch provider silently resets to openai default even when not configured, causing Dirty: yes and vector search to pause permanently due to index identity mismatch where expected model is an empty string.

Steps to reproduce

Upgrade from 2026.5.26 to 2026.6.1 via openclaw update
Run openclaw memory status
Observe provider is openai and Dirty: yes
Set provider to google: openclaw config set agents.defaults.memorySearch.provider google
Restart gateway: systemctl --user restart openclaw-gateway
Run openclaw memory status again
Observe Dirty: yes persists with indexIdentity.reason: "index was built for model gemini-embedding-001, expected "

Expected behavior

memorySearch provider should preserve previously configured value across upgrades, or default to auto which correctly selected gemini-embedding-001 in 2026.5.26. Index identity check should not compare against an empty string.

Actual behavior

Provider: openai (requested: openai)
Dirty: yes
indexIdentity:
status: mismatched
reason: "index was built for model gemini-embedding-001, expected "
providerState:
mode: pending
requestedProvider: google
Vector search: paused until memory is rebuilt
expected field is empty string — permanent mismatch regardless of provider setting.

OpenClaw version

2026.6.1 (2e08f0f)

Operating system

Rocky Linux, Node v24.14.1

Install method

npm global (openclaw update)

Model

google/gemini-3-flash-preview (main model), memorySearch previously using auto provider → gemini-embedding-001

Provider / routing chain

Google AI Studio API key → OpenClaw Gateway → memorySearch (previously auto → gemini-embedding-001, now broken)

Additional provider/model setup details

"memorySearch": {
"extraPaths": [
"/path/to/memory/"
],
"enabled": true,
"remote": {
"batch": {
"enabled": false
}
},
"provider": "google"
}

In 2026.5.26, no provider field was set — auto was the default and correctly resolved to gemini-embedding-001.

Logs, screenshots, and evidence

memory status --json key fields:
json"provider": "google",
"requestedProvider": "google",
"dirty": true,
"indexIdentity": {
  "status": "mismatched",
  "reason": "index was built for model gemini-embedding-001, expected "
},
"providerState": {
  "mode": "pending",
  "requestedProvider": "google"
},
"fts": {
  "enabled": true,
  "available": true
},
"vector": {
  "enabled": true,
  "dims": 3072
}

Impact and severity
Affected: All users who previously used Google/auto as memorySearch provider and upgrade to 2026.6.1
Severity: High — vector search completely unavailable after upgrade, index permanently dirty with no recovery path using google provider
Frequency: Always (100% reproducible after upgrade)
Consequence: Semantic memory recall degraded to FTS-only; vector similarity search unavailable indefinitely

Additional information
Last known good version: 2026.5.26 (provider: auto → correctly selected gemini-embedding-001, Dirty: no).
First known bad version: 2026.5.28 (regression first observed).
Still present in 2026.6.1.
Related issue: #90786 (google provider not recognized by --index/--deep).
Workaround: None available for users without OpenAI API key.

Impact and severity

Affected: All users who previously used Google/auto as memorySearch provider and upgrade to 2026.6.1
Severity: High — vector search completely unavailable after upgrade, index permanently dirty with no recovery path using google provider
Frequency: Always (100% reproducible after upgrade)
Consequence: Semantic memory recall degraded to FTS-only; vector similarity search unavailable indefinitely

Additional information

Last known good version: 2026.5.26 (provider: auto → correctly selected gemini-embedding-001, Dirty: no).
First known bad version: 2026.5.28 (regression first observed).
Still present in 2026.6.1.
Related issue: #90786 (google provider not recognized by --index/--deep).
Workaround: None available for users without OpenAI API key.