Skip to navigation
Configure Agents

Configure Memory Search

View as Markdown

Configure OpenClaw memory search to use an embedding model from a host Ollama container. The embedding provider is separate from your chat-model provider.

Prepare the Embedding Server

Before you continue, ensure that a separate host Ollama container is reachable on port 8000 and has the qwen3-embedding:0.6b model available. This procedure does not start the container or pull the model. Port 8000 must be unused before you start that container; do not replace an existing vLLM, NIM, or other inference service. Keep the managed Ollama daemon and authenticated proxy on their existing ports, normally 11434 and 11435.

The sandbox reaches the container through http://host.openshell.internal:8000. Bind its published port to the Docker bridge or another reviewed interface reachable from the sandbox. Allow access only from that local container network; Ollama endpoints are commonly unauthenticated. Do not change the managed Ollama daemon’s loopback binding to follow this example.

Permit the Bridge Route

The local-inference policy preset permits the sandbox to reach the host bridge. If NEMOCLAW_VLLM_PORT differs from 8000, the preset replaces its 8000 route with that port. In that case, add a custom preset for host.openshell.internal:8000 that permits the embedding server’s required GET and POST routes before running the probe. Apply the preset, then verify the Ollama API from the sandbox.

nemoclaw my-assistant policy add local-inference --yes
nemoclaw my-assistant exec -- curl -fsS http://host.openshell.internal:8000/api/tags

Continue only when the returned JSON models array contains an entry whose name is qwen3-embedding:0.6b.

Configure the Embedding Provider

Define an Ollama provider, enable the native plugin, and select the memory provider and model. OpenClaw 2026.9.1 uses memory.search for shared memory settings. Listing an embedding model under models.providers does not select it for memory search. The following settings apply to every agent without a per-agent override. To configure only main, use agents.entries.main.memory.search instead of memory.search in these commands. Existing per-agent overrides take precedence; update them if they disable search or select another provider or model.

nemoclaw my-assistant exec -- \
  openclaw config set models.providers.ollama-mem \
  '{"api":"ollama","baseUrl":"http://host.openshell.internal:8000","apiKey":"ollama-local","models":[{"id":"qwen3-embedding:0.6b","name":"Qwen3 Embedding 0.6B"}]}' \
  --strict-json \
  --merge
nemoclaw my-assistant exec -- openclaw plugins enable ollama
nemoclaw my-assistant exec -- \
  openclaw config set memory.search.enabled true --strict-json
nemoclaw my-assistant exec -- \
  openclaw config set memory.search.provider ollama-mem
nemoclaw my-assistant exec -- \
  openclaw config set memory.search.model qwen3-embedding:0.6b
nemoclaw my-assistant exec -- \
  openclaw config set memory.search.fallback none
nemoclaw my-assistant gateway restart

ollama-local is a non-secret placeholder; it does not authenticate an otherwise protected server. Disabling fallback keeps verification on the selected embedding provider. The restart activates the plugin and memory settings. OpenClaw owns the native configuration after first launch. Restart and reconnect preserve the current native changes. Rebuild transfers the native configuration captured immediately before replacement. Before gateway startup, NemoClaw refreshes the scoped inference credential handle when its endpoint and credential key match the native NVIDIA or OpenRouter provider. Refer to Understand Runtime Changes for the reconciliation rules. To roll back to earlier memory-search settings, reapply them explicitly from an independent trusted record; the retired selective snapshot restore flow is not available. The local-inference OpenShell policy preset controls access to the private host bridge URL.

Rebuild and Verify the Index

An existing index can report an identity mismatch after its embedding provider or model changes. Reindexing replaces the derived index and sends indexed memory text to the configured embedding server. It preserves the source memory files. Rebuild the index, then probe the embedding provider and vector store:

nemoclaw my-assistant exec -- openclaw memory index --agent main --force
nemoclaw my-assistant exec -- openclaw memory status --agent main --deep --json

For each selected agent, confirm that embeddingProbe.ok and status.vector.semanticAvailable are true. Confirm that status.model is qwen3-embedding:0.6b and status.vector.index.state is complete for a nonempty corpus. An empty memory corpus produces no search results; add a memory before testing retrieval.

Managed startup exports SQLITE_TMPDIR to standalone OpenClaw commands. Search for a fact from an existing memory, using different wording to check semantic retrieval.

nemoclaw my-assistant exec -- openclaw memory search --agent main \
  --query 'a description of a fact saved in your memory' --json

Confirm that the returned file and snippet contain the expected fact. Results report vectorScore and textScore; a positive vector score with a zero text score demonstrates retrieval without keyword contribution. Replace main with your intended agent ID in each memory command. Omitting --agent from index or status processes every configured agent.