ollama-memory-embeddings_skill

Overview

This skill configures OpenClaw memory search to use Ollama as the embeddings server via its OpenAI-compatible /v1/embeddings endpoint. It replaces the default local node-llama-cpp embedding generation with selectable Ollama embedding models and offers optional import of local GGUF embedding files into Ollama. The installer is idempotent, creates safe backups, and includes verification, drift enforcement, and optional auto-heal tooling.

How this skill works

The installer verifies Ollama is reachable, presents interactive model selection (embeddinggemma, nomic-embed-text, all-minilm, mxbai-embed-large), and normalizes the chosen model name with a :latest tag. It updates only the keys this skill owns in OpenClaw config (provider, model, remote.baseUrl, remote.apiKey), performs post-write JSON validation, and can restart the OpenClaw gateway and optionally reindex memory. It can detect local embedding GGUFs and import them into Ollama, and it includes enforce and watchdog scripts to detect and heal configuration drift.

When to use it

• You want OpenClaw to generate memory embeddings via a running Ollama server instead of local node-llama-cpp.
• You need to standardize embedding generation across machines or centralize embeddings behind Ollama.
• You have local GGUF embedding files you want available through Ollama without manual model creation.
• You require automated drift detection and idempotent enforcement for memory search config.
• You plan to change embedding models and want safe rebuild/reindex guidance and tooling.

Best practices

• Back up your OpenClaw config before running; the installer already creates timestamped backups when it writes.
• Choose reindex-memory when switching embedding models to avoid mismatched vector spaces.
• Use --dry-run to preview changes in non-interactive scripts and CI.
• Opt into importing local GGUFs explicitly (--import-local-gguf yes) to avoid unintended model creation.
• Install the watchdog on systems where config drift is likely (macOS uses launchd install option).

Example use cases

• Switch an existing OpenClaw instance to use Ollama-hosted embeddings without touching chat/completions routing.
• Import a high-quality local GGUF (mxbai-embed-large) into Ollama to serve consistent embeddings across deployments.
• Automate configuration drift healing on macOS by installing the launchd watchdog to restore desired memorySearch settings.
• Run non-interactive installs in containers or automated setups with specified model and reindexing behavior.
• Verify an Ollama embedding endpoint is functioning by running the included two-step verification and verbose diagnostics.