Bug type
Tool/runtime regression: CLI search works, agent tool fails.
Summary
The memory_search tool invoked from the agent returns disabled: true with "index metadata is missing", even though the CLI reports the memory index as healthy and CLI search returns results.
This happens after rebuilding the index and restarting the gateway.
Environment
- OpenClaw: 2026.6.1
- Install: npm global inside Docker container
- Deploy: Docker on VPS
- Memory plugin: memory-core
- Memory backend: builtin SQLite + sqlite-vec
- Embedding provider: OpenAI
- Embedding model: text-embedding-3-small
- Agent: main
Both CLI paths report the same version:
/usr/local/bin/openclaw --version
/data/.npm-global/bin/openclaw --version
Both return:
OpenClaw 2026.6.1 (2e08f0f)
Steps to reproduce
- Rebuild memory index:
openclaw memory index --force
-
Restart the gateway/container.
-
Check index status:
openclaw memory status --index --agent main
- Run CLI search:
openclaw memory search --agent main --query "proyecto wordpress agentes"
- Invoke
memory_search from the agent tool chain with a similar query.
Expected behavior
memory_search should use the same ready index and return results.
Actual behavior
CLI status reports the index is healthy:
Memory index complete.
Memory Search (main)
Provider: openai (requested: openai)
Model: text-embedding-3-small
Sources: memory
Indexed: 18/18 files · 58 chunks
Dirty: no
Embeddings: ready
Vector store: ready
Semantic vectors: ready
Vector dims: 1536
FTS: ready
CLI search works and returns results.
But agent tool invocation fails with:
{
"disabled": true,
"error": "index metadata is missing"
}
In some cases the tool also reports that the index was built with different embedding provider/model/settings, despite CLI status showing openai/text-embedding-3-small.
Workaround
Use the CLI explicitly with the agent id:
openclaw memory search --agent main --query "<query>"
This works while the memory_search tool fails.
Additional context
This may be either:
- the
memory_search tool checking metadata differently than openclaw memory index --force writes it, or
- the tool resolving a different/default agent index than
main.
The CLI default search behavior also appears sensitive to agent selection; --agent main works reliably.
Bug type
Tool/runtime regression: CLI search works, agent tool fails.
Summary
The
memory_searchtool invoked from the agent returnsdisabled: truewith"index metadata is missing", even though the CLI reports the memory index as healthy and CLI search returns results.This happens after rebuilding the index and restarting the gateway.
Environment
Both CLI paths report the same version:
Both return:
Steps to reproduce
Restart the gateway/container.
Check index status:
openclaw memory search --agent main --query "proyecto wordpress agentes"memory_searchfrom the agent tool chain with a similar query.Expected behavior
memory_searchshould use the same ready index and return results.Actual behavior
CLI status reports the index is healthy:
CLI search works and returns results.
But agent tool invocation fails with:
{ "disabled": true, "error": "index metadata is missing" }In some cases the tool also reports that the index was built with different embedding provider/model/settings, despite CLI status showing
openai/text-embedding-3-small.Workaround
Use the CLI explicitly with the agent id:
openclaw memory search --agent main --query "<query>"This works while the
memory_searchtool fails.Additional context
This may be either:
memory_searchtool checking metadata differently thanopenclaw memory index --forcewrites it, ormain.The CLI default search behavior also appears sensitive to agent selection;
--agent mainworks reliably.