OpenClaw: "Managed local embeddings are unavailable"
When a saved llama.cpp server path points to a missing executable, reindexing cannot fix it. Restore the provider on the Gateway host, then rebuild memory search.
“Managed local embeddings are unavailable” can mean OpenClaw has a configured llama.cpp server path but the executable is missing on this machine. Rebuilding the memory index cannot start a file that is not there. Check the embedding provider and its configured command, then run the llama.cpp provider’s guided setup on the Gateway host. Restore local embeddings before rebuilding the index.
The configuration survived; the executable did not
Local embeddings turn text into the vectors used for semantic memory search. They are separate from the model answering your chat. A working conversation therefore does not prove that local memory search works.
The missing-server behavior arrived in 2026.9.2 and is present through 2026.9.5: when localService.command is configured, the managed provider starts that path without first installing a missing server. The upstream report records the version boundary and the configured-path failure.
That matters when restoring or moving an installation. A copied configuration can point into a tools directory that was not copied, or at an absolute path that belongs to the old machine. The config is present, but the dependency it names is not.
Check which part is unavailable
openclaw memory status --deep --agent main
openclaw config get models.providers.llama-cpp.localService.commandReplace main with the affected agent’s ID. Check the returned command path on the Gateway’s computer, inside its container if it runs in one. Installing a server on your laptop does not repair a remote Gateway.
The memory diagnostic reference distinguishes embedding readiness from vector-store readiness. A vector-store loading warning is not the same as an unavailable embedding provider. If the executable exists, read the actual startup error rather than assuming this missing-file case.
Run the provider’s setup on this host
The local-service documentation says to run llama.cpp setup on each Gateway host, rather than copying the generated command path between machines. Its setup chooses and verifies the matching platform build. Follow the llama.cpp provider’s guided setup and select the managed local server or local-embedding choice appropriate to your version.
On 2026.9.4, the repair hint is:
openclaw models auth login --provider llama-cpp --method localThat path requires a terminal and interactive consent. It is not interchangeable with connecting an already-running external llama-server. The missing-file report remained applicable through 2026.9.5; repeating a non-interactive setup command that does not support the provider is not a repair.
Then rebuild and search
openclaw memory status --deep --agent main
openclaw memory index --force --agent mainProceed with indexing once the provider is available, then test a real memory search. Do not delete the agent database: it also holds other agent state. If the pause instead names a changed chunking implementation after an update, that has a different recovery path, described in the update guide. Other provider and index failures belong in the broader memory-search troubleshooting guide.