How does OpenClaw find memories by meaning rather than exact words? Vector search. The sqlite-vec extension adds vector columns to SQLite tables. Each memory chunk is embedded using EmbeddingGemma, a 300M parameter model that runs locally.
When you ask a question, OpenClaw embeds your query into the same vector space and finds the closest matches using cosine similarity. The model downloads automatically on first use (about 600MB). After that, all embedding happens offline with no API calls. This means your memory data never leaves your machine.